🛰️ Frontier Digest Log

[chrono-meta]

Standing durable store for the self-evolution routine (knowledge/shared/harness-core/self_evolution_routine.md).

• The Daily routine posts each day's frontier-digest here as a comment (HN + arxiv signals, each with a source URL + one-line FH relevance).
• The Weekly routine reads the last 7 days of comments on this issue, runs persona-innovator Mode F, and opens at most one draft PR (HITL merge).

This issue is the ephemeral-safe cross-run memory: routine sessions are fresh clones, so signals live here on GitHub, never in the (gitignored) tracks/** tree.

Setup: paste this issue's number into the {ISSUE} placeholder of both routine prompts (doc §7) before saving the routines at claude.ai/code/routines.

[chrono-meta]

🔭 FH Frontier Digest — 2026-06-16

🔑 Engine: WebSearch mode (no ANTHROPIC_API_KEY) · in-context synthesis

📡 Collected: HN ~10 items · arxiv 8 items · TLDR/Batch (meta only)

---

This Week's Frontier Highlights

• SkillRevise: Trace-Conditioned Skill Revision (arxiv:2606.01139) — LLM-authored agent skills revised via execution traces under a shared verifier-driven harness; GPT-5.5 task success raised from 36% → 61% in 3 revision rounds. Companion paper SkillAdaptor (2606.01311) shows self-adaptation from trajectories without manual intervention.

  FH relevance: validates the predict-verify loop at the heart of FH's edit-manifest + harvest-loop cycle; trace-conditioned revision is a concrete mechanism to add to the skill improvement cadence alongside steel-quench.

• Self-Healing Agentic Orchestrators (arxiv:2606.01416) — Classifies orchestration failure modes (tool timeouts, malformed args, stale context, retry loops, unverified intermediate outputs) and maps each to targeted bounded recovery actions; verifies recovered trajectories before continuing.

  FH relevance: mcp-circuit-breaker could adopt this failure-class taxonomy as structured detection criteria; the bounded-recovery pattern complements verify-bidirectional's baseline-update circuit.

• Anthropic Three-Agent Harness: Planner / Generator / Evaluator (NewClawTimes · InfoQ) — Experimental data: without Evaluator arm → 20 min / $9 / broken output; with full P→G→E loop → 6 h / $200 / functionally correct. Agents coordinate via sprint contracts (shared file on disk).

  FH relevance: FH's Diagnostic Triad (harness-doctor + context-doctor + sim-conductor) is architecturally isomorphic with P/G/E — this result grounds the value of FH's verify step with empirical cost-quality data. "Sprint contracts" = FH's Done-When definitions.

• Malicious Skill Infiltration in Agent Marketplaces (HN:48015397 · HN:48265332) — 1,200 malicious skills infiltrated the OpenClaw agent marketplace; researchers subsequently catalogued 6,487 malicious agent tools that VirusTotal cannot detect. Proposed mitigation: a hypervisor-type authorization layer outside the sandbox.

  FH relevance: directly strengthens the case for FH's skill pre-commit gate and marketplace-gate; consider adding provenance verification (signed commit + origin repo check) to the skill onboarding flow.

• Less Context, Better Agents (arxiv:2606.10209) — Pruning agent context to the last 5 tool calls achieved 79% task completion with a 63.9% token reduction; adding automated summarization raised completion to 91.6%. Conclusion: careful context engineering outperforms raw context-window expansion.

  FH relevance: directly validates FH's context-doctor and .claudeignore design philosophy; suggests adding a tool-call window pruning heuristic as a concrete context-doctor recommendation.

• Agentic Environment Engineering Survey (arxiv:2606.12191) — Survey of agent evolution across 4 axes: memory-centric, orchestration-centric, trajectory-centric, and exploration-centric. Complements the earlier Externalization survey (arxiv:2604.08224) covering memory, skills, protocols, and harness engineering as a unified framing.

  FH relevance: trajectory-centric axis maps directly to FH's harvest-loop; the 4-axis evolution taxonomy provides external vocabulary for FH's 6-axis framework when writing documentation for external audiences.

---

⚠️ Warning Signals

• The $200/6h vs $9/20min framing in the Anthropic harness report risks anchoring teams on raw cost as a quality proxy. FH's Done-When + gate structure is the cheaper path to the same reliability — distinguish "evaluator loop" (architectural) from "more compute" (blunt).
• Malicious skill marketplace contamination is accelerating faster than detection tooling. FH's current pre-commit gate is form-check only (provenance is explicitly a residual per CLAUDE.md); this is the right honest scope now, but the gap between form-check and full provenance verification is widening.

---

Autonomous daily scan · WebSearch mode · No changes committed

---

Generated by Claude Code

[chrono-meta]

🔭 FH Frontier Digest — 2026-06-17

🔑 Engine: WebSearch mode (no ANTHROPIC_API_KEY in sandbox)
📡 Collected: HN ~8 items · arxiv ~10 items | Autonomous collection arm only

---

This Week's Frontier Highlights

1. Self-improving harness loops are now measured, not theoretical
Three independent arxiv papers (May–June 2026) describe agents that autonomously improve their own operating harness without human engineers: Self-Harness (weakness mining → harness proposals → validation, +15–52% on Terminal-Bench 2.0), SIA (+25% on LawBench), and Retrospective Harness Optimization / RHO (single round lifts SWE-Bench Pro from 59% → 78% via self-preference over trajectory rollouts).
Sources: arxiv:2605.27276 · arxiv:2606.05922 · arxiv:2605.09998

Relevance to FH: FH's 4-axis auto-gate (Mode D) is architecturally identical to the Self-Harness three-stage loop (regression_guard ≈ weakness mining, steel-quench ≈ harness proposals, phantom-quench ≈ validation). The RHO paper directly validates FH's edit-manifest predict-verify design — strengthens the case for harvest-loop closing the Axis-4 measurement loop.

2. Anthropic publishes "Effective Harnesses for Long-Running Agents"
Vendor-level engineering guidance covering 30 lifecycle hook events, parallel sub-agent orchestration, and token-budget management in production. Emerged alongside the Claude Code 2.1.88 source-map analysis that surfaced 12 reusable agentic harness design patterns.
Sources: anthropic.com/engineering · Epsilla blog

Relevance to FH: Direct vendor guidance — cross-check the 12 patterns against FH's templates/ and Full-Harness Mode §6 install list. Any gap is a candidate for the next Full-Harness template update. The 30-hook event model is substantially richer than FH's current SessionStart/Stop hooks.

3. "Externalization in LLM Agents: Memory, Skills, Protocols and Harness Engineering" survey
Four-paradigm taxonomy: Monolithic Context → Context with Retrieval → Hierarchical Memory + Orchestration → Adaptive Memory Systems. Published April 2026 as a unified frame for the field.
Source: arxiv:2604.08224

Relevance to FH: FH's skill bus + MEMORY.md maps to the "Hierarchical Memory and Orchestration" paradigm. Gap signal: FH memory is currently prose-driven (MEMORY.md advisory, no mechanical retrieval). The survey's Adaptive Memory tier is where the frontier is headed — a UAP + episodic retrieval integration is the natural next step.

---

FH Immediate Application Candidates

• Adopt Self-Harness weakness-mining step into harvest-loop: When weekly audit enumerates defects, auto-generate harness-improvement proposals (harness-doctor output → candidate patches) — closing the loop that currently requires manual human triage.
• Cross-check Anthropic's 12 agentic patterns against FH templates/: Run as a /sister-asset scan; any unrepresented pattern is a fh_signal candidate for Full-Harness Mode §6.
• Evaluate RHO-style trajectory rollout for edit-manifest: Axis-4 currently records predicted impact. RHO shows that comparing trajectory rollouts (predicted vs actual across sessions) drives measurable improvement — harvest-loop Step 5 could add this comparison pass.

---

Warning Signals

• "Code as Agent Harness" push (arxiv:2605.18747) argues prose-driven rules are the weakest substrate for agent governance. FH's prose-heavy CLAUDE.md ruleset is a structural vulnerability by this framing. Mitigating factor: FH's pre-commit hook already moves critical gates to code — but the majority of FH rules remain salience-dependent prose.
• Harness complexity creep: n8n blog notes the field is re-learning what "agent tools" even are in 2026 — fragmentation risk. FH's "simpler over time" guard and harness-doctor 30d cadence are the right countermeasure; worth re-running harness-doctor given the volume of 2026 additions.

---

Autonomous daily scan · no commit · no PR · collection only

---

Generated by Claude Code

[chrono-meta]

🔭 FH Frontier Digest — 2026-06-18

🔑 Engine: WebSearch mode (no ANTHROPIC_API_KEY; in-context synthesis from HN + arxiv results)

---

This Week's Frontier Highlights

• Harness > Model (HN #46988596) — "Improving 15 LLMs at Coding in One Afternoon. Only the Harness Changed" — swapping only the harness moved a coding agent from Top 30 to Top 5 on Terminal Bench 2.0.

  • Source: https://news.ycombinator.com/item?id=46988596
  • Relevance to FH: Validates FH's core thesis; harness-doctor + regression_guard.sh are first-class differentiation levers, not overhead.

• ACE: Agentic Context Engineering (ICLR 2026, arXiv:2510.04618) — Treats session contexts as evolving playbooks that accumulate/refine strategies through generation → reflection → curation. Prevents "context collapse" (iterative rewriting erodes details). +10.6% on agent benchmarks, +17.1% on AppWorld.

  • Source: https://arxiv.org/abs/2510.04618
  • Relevance to FH: Direct analogue of FH session cards + compounding loop. "Context collapse" ≈ FH's card-update debt bug (incomplete card after context compression). ACE's curation phase maps to harvest-loop Step 0-b.

• Claude Code v2.1.172 — recursive sub-agent spawning (5 levels deep) — Released 2026-06-10; sub-agents can now spawn their own sub-agents with nesting up to 5 levels. Planning agents dispatch specialist agents that themselves manage parallel workloads.

  • Source: https://smartscope.blog/en/generative-ai/claude/claude-code-best-practices-advanced-2026/
  • Relevance to FH: FH's parallel-dispatch pattern gains new surface area; stale-context bug exposure expands with depth. FH's "Agent View pre-read" guard and session card protocol need explicit depth-governance guidance.

• "Context engineering is the fix(meta-validation): Wave 3 — table count, CI agent detect, path neutralize + external network artifacts #1 job for engineers building AI agents" (arXiv:2603.09619) — 2026 industry consensus: most agent failures are context failures, not model failures. Hierarchical memory architectures (short-term / working / long-term) are the main 2026 focus area.

  • Source: https://arxiv.org/pdf/2603.09619
  • Relevance to FH: Reinforces context-doctor + .claudeignore as load-bearing FH assets. May warrant a dedicated context-drift detection axis in harness-doctor L2+.

• 65% of enterprise AI failures = Harness Defects (Context Drift · Schema Misalignment · State Degradation) — Augment Code / industry survey; "Teams shipping AI agents successfully aren't the ones with the best models — they're the ones with the best evaluation infrastructure."

  • Source: https://www.augmentcode.com/guides/harness-engineering-ai-coding-agents
  • Relevance to FH: harness-doctor L1–L4 axes map well to first two defect classes; state-degradation detection is an open gap in current FH toolset — candidate for L3 prescription.

• Counter-signal: "Building agents without harness engineering" (HN #48493447) — Growing practitioner pushback on harness overhead; minimalist agent patterns gaining traction alongside the structured camp.

  • Source: https://news.ycombinator.com/item?id=48493447
  • Relevance to FH: FH's "simpler over time / complexity = warning signal" guard is the correct hedge. Steel-quench Wave-1 angle fix(meta-validation): Wave 3 — table count, CI agent detect, path neutralize + external network artifacts #1 ("is there no simpler alternative?") should stay a first-pass gate.

---

FH Immediate Application Candidates

1. harness-doctor L3 — add state-degradation detection prescription — the "State Degradation" defect class (session state that silently drifts across turns) is named in enterprise surveys as a top-3 failure mode but has no dedicated FH diagnostic today.
2. context-doctor upgrade — ACE's "context collapse" framing + hierarchical memory taxonomy could sharpen context-doctor's diagnosis vocabulary (currently focused on token footprint, not content-drift).
3. Depth-governance note for parallel-dispatch pattern — With 5-level recursive sub-agents now live, FH's agent-dispatch guidance (CLAUDE.md §Agent Dispatch Operation) needs a one-line depth ceiling recommendation and a stale-context-at-depth guard.

Warning Signals

• The "harness engineering" term is now widely used outside FH — ecosystem convergence is healthy, but FH assets that use the term without external links may appear parochial. A quarterly phantom-quench pass on external terminology drift is recommended.
• Recursive sub-agent nesting at depth creates context-isolation gaps not covered by existing FH pre-read guards — monitor for session-card stale-content bug reports at depth ≥ 3.

---

📊 Collected: HN ~10 items · arxiv 4 papers · WebSearch synthesis (no API key) | Autonomous daily run — no operator online

---

Generated by Claude Code

[chrono-meta]

Frontier Digest — 2026-06-19

Autonomous collection run · WebSearch mode · HackerNews + arxiv · 5 signals

---

• System scaling now outranks model scaling as the primary AI bottleneck — arxiv 2605.26112 "From Model Scaling to System Scaling" frames the harness (context governance · trustworthy memory · dynamic skill routing) as the first-class object of design, not the model itself. Claude Code is cited as a production exemplar. | FH relevance: direct external validation of the 6-axis framework; the three named bottlenecks map almost 1-to-1 to FH Axes 2/3/4 — strengthens the case for context-doctor and hub_compounding_loop as core, not optional, assets.

• Trace-guided harness repair improves held-out test performance 15–50% over human-designed harnesses — arxiv 2606.06324 "From Failed Trajectories to Reliable LLM Agents" introduces HarnessFix: raw execution traces → HTIR (Harness-aware Trace Intermediate Representation) → step-level root-cause attribution → automated repair. Also beats self-evolution baselines. | FH relevance: FH's regression_guard (Axis 1) + harness-doctor cover adjacent ground; the HTIR formalism (structured provenance per step) is a concrete technique FH's prompt-regression and harness-doctor could absorb to move from heuristic diagnosis to trace-grounded diagnosis.

• Anthropic internal evals: context editing alone = +29% perf lift; combined with memory tool = +39% — reported in the 2026 Context Engineering Agent Reliability Playbook (digitalapplied.com) citing Anthropic data. Most production teams treat context as an afterthought and only discover failures in long sessions. | FH relevance: quantitative anchor for context-doctor's value proposition; the "+10pp gain from adding a memory tool" supports FH's two-layer storage principle (tracks/ + companion store) as a measurable engineering decision, not just hygiene.

• RL for multi-agent orchestration decomposes into 5 sub-decisions — "when to stop" has no RL method yet — arxiv 2605.02801 models orchestration as temporal interaction graphs with 8 reward families; spawning/delegation/aggregation are covered by existing RL methods, but the stopping decision has no principled method as of May 2026. | FH relevance: FH's mcp-circuit-breaker + goal-quench guard the stopping surface heuristically; this gap is an opportunity to adopt/watch for RL-based stopping methods when they emerge; the 5-decision decomposition is also a useful lens for auditing agent-composer's dispatch logic.

• Paired adversarial evaluation (human-in-the-bridge) outperforms isolated benchmarks for regression testing — arxiv 2606.16871 "Human-on-the-Bridge" shows that comparing two agent versions under identical adversarial pressure is more reliable than benchmark scores for detecting regressions; introduces ProofAgent Harness for auditable multi-juror evaluation. | FH relevance: steel-quench's Wave-1 adversarial angle is structurally aligned with this finding; the "multi-juror scoring + disagreement resolution" pattern is a concrete upgrade candidate for steel-quench's judged conditions (currently single-pass adversarial review → add a second independent pass before a judged verdict is final).

---

Sources: arxiv 2605.26112 · arxiv 2606.06324 · Context Engineering Playbook · arxiv 2605.02801 · arxiv 2606.16871

---

Generated by Claude Code

[chrono-meta]

🔭 FH Frontier Digest — 2026-06-20

🔑 Engine: WebSearch mode (no ANTHROPIC_API_KEY; curl blocked by egress policy → WebSearch fallback per SKILL.md §Simplification Guards)

📡 Collected: HN blocked (egress) · arxiv via WebSearch · supplementary sources via WebSearch | 6 signals synthesized

---

This Week's Frontier Highlights

• Self-Harness: Harnesses That Improve Themselves (arXiv:2606.09498, Jun 8 2026)
  LLM-based agents autonomously optimize their own harness via a three-stage loop — Weakness Mining (failure pattern extraction from execution traces) → Harness Proposal (minimal targeted edits) → Proposal Validation (regression-gated acceptance). Achieved 33–60% gains on Terminal-Bench 2.0 without human engineers or stronger external models.
  — Relevance to FH: FH's 4-axis auto-gate (regression_guard → steel-quench → phantom-quench → edit-manifest) is structurally isomorphic with this loop; this paper empirically validates the "autonomous self-verification before commit" architecture. The "regression-gated acceptance" step directly mirrors Axis 1.
  📎 https://arxiv.org/abs/2606.09498

• Less Context, Better Agents: Efficient Context Engineering for Long-Horizon Tool-Using LLM Agents (arXiv:2606.10209, Jun 2026)
  Verbose MCP tool responses cause context overflow and stale-state errors in enterprise workflows. Selective retention of recent tool interactions + compact summarization achieves 91.6% task completion at 553K tokens vs full-history retention degrading at ~32K tokens.
  — Relevance to FH: Empirically validates FH's /context-doctor skill and token-budget-gate; the paper's "quality over quantity" finding aligns directly with FH's simplification guard ("simpler over time") and tiered execution settings (S/M/L/XL). Specifically relevant to FH's MCP tool gating — verbosity is a measured risk, not a hypothetical.
  📎 https://arxiv.org/abs/2606.10209

• AdaptOrch: Task-Adaptive Multi-Agent Orchestration in the Era of LLM Performance Convergence (arXiv:2602.16873, Feb 2026)
  As major frontier LLMs (GPT-4o, Claude 3.5 Sonnet, Gemini 2.0, Llama 3.3 70B) converge within 2–5% on MMLU/HumanEval/MATH, orchestration topology — not model selection — is the dominant system-level performance lever. AdaptOrch dynamically selects among parallel, sequential, hierarchical, and hybrid topologies based on task dependency graph (DAG) properties.
  — Relevance to FH: FH's /agent-composer dispatches parallel agents but applies no topology selection logic. The DAG-based routing framing is a concrete improvement candidate — when 2+ agents are dispatched, the dependency graph shape should guide topology, not just "are tasks independent?" heuristics.
  📎 https://arxiv.org/abs/2602.16873

• Adaptive Auto-Harness: Sustained Self-Improvement for Open-Ended Task Streams (arXiv:2606.01770, Jun 2026)
  A continuous adaptation loop for deployed agentic systems across streaming task environments — addresses the failure mode where a harness optimized for one task distribution degrades on distribution shift, without requiring human-triggered reset.
  — Relevance to FH: Structural parallel with FH's Operational Adaptation Loop (UAP in tracks/_meta/user_adaptation_profile.md). The "open-ended task stream" framing extends UAP's per-user skill-proposal adaptation toward per-task-distribution adaptation — a potential UAP v2 generalization trigger if the 2+ session recurrence threshold fires.
  📎 https://arxiv.org/abs/2606.01770

• MCP Ecosystem: 9,652 Servers in Official Registry (May 2026); Stateless Spec Targeting June 2026
  As of May 24 2026, the official MCP registry holds 9,652 latest server records. 41% of surveyed software organizations are in limited or broad production with MCP servers. The next MCP spec release (June 2026) targets stateless transport while supporting stateful apps. Enterprise Okta integration now provisions MCP connectors zero-touch at first login across Claude chat, Claude Code, and Cowork.
  — Relevance to FH: FH's mcp_tool_gating.md (templates rule: never trust server annotations) becomes more load-bearing, not less, as the surface area grows 10x. The "all-None hints" measured failure (2026-06-11, PR docs(ax-coach): Advanced patterns from QASP session #109 context) is now the industry norm — FH's ask/allow gating is validated practice, not conservative overhead.
  📎 https://chatforest.com/guides/mcp-ecosystem-2026-state-of-the-standard/

• TokenPilot: Cache-Efficient Context Management for LLM Agents (arXiv:2606.17016, Jun 2026)
  Cache-aware context management that exploits KV-cache reuse patterns across agent turns to reduce inference cost without accuracy loss — complements summarization-based approaches (cf. arXiv:2606.10209 above) by operating at the cache layer rather than the content layer.
  — Relevance to FH: Secondary signal for /token-budget-gate; if FH ever moves to multi-turn agent loops with shared KV context (e.g., extended /goal runs), cache-aware context ordering is a latency + cost lever worth tracking.
  📎 https://arxiv.org/abs/2606.17016

---

FH Immediate Application Candidates

1. 4-axis gate ↔ Self-Harness alignment note — add one-line cross-reference in harness_6axis_framework.md §Axis 5 pointing to arXiv:2606.09498 as external validation of the regression-gated commit pattern. No new infra — annotation only.
2. agent-composer topology hint — surface AdaptOrch's DAG topology taxonomy (parallel / sequential / hierarchical / hybrid) as a decision aid in /agent-composer Step 1 when 2+ agents are dispatched. Candidate for a fh_signal → /field-harvest pass.
3. context-doctor verbosity warning — add MCP tool response verbosity as an explicit detection signal in /context-doctor (mirrors arXiv:2606.10209 enterprise finding); currently context-doctor detects large files, not large tool payloads.

Warning Signals

• The "Self-Harness" and "Adaptive Auto-Harness" cluster (3+ papers in June 2026 alone) shows rapid convergence on autonomous harness optimization — risk of FH reinventing what's becoming a standard research primitive. Check arXiv:2606.09498 + 2606.01770 against FH's 4-axis gate scope before building new self-improvement machinery. No new build recommended until cross-audit completes.
• MCP ecosystem size (9.6K servers) creates a broader attack surface for the "trust server annotations" failure mode FH already measured. mcp_tool_gating.md's ask/allow table needs a sweep cadence, not just a mount-time fill.

---

Autonomous daily scan · WebSearch mode · No commits, no PRs · Next operator-intake cadence: 7-day frontier-digest schedule

---

Generated by Claude Code

[chrono-meta]

🔭 FH Frontier Digest — 2026-06-21

🔑 Engine: WebSearch mode (curl blocked in sandbox; no ANTHROPIC_API_KEY — synthesis performed in-context by session model)

📡 Collection: HN discussion · Anthropic blog · Linux Foundation AAIF · O'Reilly Radar · LangChain SoAE survey

---

Frontier Signals

• MCP becomes a neutral Linux Foundation open standard via AAIF — Anthropic donated the Model Context Protocol to the Agentic AI Foundation (Anthropic + OpenAI + Block co-founders; AWS, Google, Microsoft, Cloudflare platinum members). Now 6,000+ app integrations, 75+ Claude connectors; Enterprise-Managed Authorization extension shipped.
  Source: Anthropic — Donating MCP & establishing AAIF · Linux Foundation AAIF press release
  Relevance to FH: mcp_tool_gating.md's "server annotations are unreliable" finding remains valid, but AAIF membership may stratify trust tiers; the name-keyed ask/allow table should track AAIF-governed servers as a distinct tier.

• Claude Code "A harness for every task": dynamic workflows (June 2, 2026) — Claude Opus 4.8 can now write its own multi-agent harness on the fly (classifier→router and splitter→synthesizer patterns). Token-heavy; recommended for complex, high-value, or adversarial tasks only.
  Source: Anthropic blog — A harness for every task · Lushbinary guide
  Relevance to FH: Tier-0 platform capability directly overlaps agent-composer and goal-quench. FH should route simple orchestration to dynamic workflows and reserve FH meta-orchestration for governance-aware / gate-enforced flows — no-reinvention principle at stake.

• "Context engineering" is the 2026 production discipline — O'Reilly Agents Stack (2026 edition) names context engineering as the successor to prompt engineering. Context rot is real: performance degrades around 32K tokens even with 1M context windows. Memory is now a 3-tier primitive (in-context / retrieved / persistent); minimum high-signal tokens is the production pattern.
  Source: O'Reilly — The AI Agents Stack (2026 Edition)
  Relevance to FH: Validates and names FH's existing context-doctor / .claudeignore approach. Update context-doctor triggers to include "context engineering" as a synonym for external discoverability; the 32K rot threshold is a candidate concrete gate value.

• AGENTS.md joins Linux Foundation AAIF alongside MCP — OpenAI contributed AGENTS.md (agent instruction standard) and Block contributed goose to the AAIF; these are now Linux Foundation projects on a convergence path with MCP.
  Source: Linux Foundation AAIF announcement · GitHub Blog — MCP joins Linux Foundation
  Relevance to FH: FH ships AGENTS.md as the Codex/external-entry point. An alignment check against the emerging AAIF AGENTS.md community spec is warranted before the next npm publish (CLAUDE.md §④-b lockstep applies).

• 57% of practitioners have agents in production (LangChain State of Agent Engineering 2026) — LangGraph is the leading production framework; reliability, determinism, and HITL are the dominant pain points. Shift from demo-ware to reliability-first validates harness engineering investment.
  Source: LangChain — State of Agent Engineering
  Relevance to FH: FH's "simpler over time" field-harness principle and HITL gates (goal-quench budget gate, 4-axis pre-commit) are aligned with production practitioner demands — this is a signal FH's governance-first angle has external resonance.

• HN top thread: "The current hype around autonomous agents, and what actually works in production" — Active practitioner debate surfacing real failure modes: hallucination under long context, verification gaps, missing human-in-the-loop checkpoints.
  Source: HN item #44623207
  Relevance to FH: Failure modes reported (context bloat, no verification gate, no adversarial check) map directly to FH's steel-quench (adversarial), sim-conductor (verification), and context-doctor (bloat) — FH's 4-axis gate is answering a pain point the HN practitioner community is actively experiencing.

---

FH Immediate Application Candidates

1. mcp_tool_gating.md — add AAIF-governed tier: AAIF membership is now a meaningful provenance signal; the existing ask/allow table should gain an AAIF-member flag as a soft-trust baseline (still name-keyed, not annotation-trusted).
2. context-doctor trigger update: add "context engineering" and "context rot" to trigger phrases in SKILL.md for external discoverability; add the ~32K token degradation threshold as a concrete gate recommendation.
3. agent-composer / goal-quench — no-reinvention alignment: document explicit routing rule: dynamic-workflows-eligible tasks → platform Tier-0; gate-aware / FH-governed flows → FH meta-orchestration. Prevents silent reinvention as dynamic workflows matures.

Warning Signals

• Dynamic workflows is token-heavy by design — FH skills that auto-invoke multi-agent runs (e.g. sim-conductor, pipeline-conductor) should flag this cost asymmetry; otherwise user perception of FH as expensive will grow as dynamic workflows becomes the ambient default.
• AAIF governance is promising but early — do not prematurely upgrade MCP server trust tiers based on AAIF membership alone; the "annotations are unreliable" finding was measured empirically and should not be retired on governance grounds alone.

---

📊 Collected: HN 1 discussion · Anthropic/AAIF announcements · O'Reilly Radar 1 article · LangChain survey · 3 supplementary sources | Autonomous daily scan — no operator online

---

Generated by Claude Code

[chrono-meta]

Frontier Digest — 2026-06-22

🔑 Engine: WebSearch mode (no ANTHROPIC_API_KEY) · Sources: HackerNews + arxiv · Autonomous daily scan

---

Signals relevant to FH skills/structure

• HarnessFix — trace-grounded harness repair (arxiv:2606.06324): Proposes HTIR (Harness-aware Trace Intermediate Representation) to diagnose which harness layer caused a failure from raw execution traces, then applies targeted repairs. Reports 15–50% improvement over initial harnesses and out-performs human-designed baselines. Explicitly references ETCLOVG layers as its diagnostic taxonomy.
  Relevance to FH: Directly validates FH's Axis-5 (verify) + Axis-6 (improve) design and the verify-bidirectional + harness-doctor loop. ETCLOVG layer labels appearing in an external paper are a cross-reference signal worth tracking.

• Self-Harness — harnesses that improve themselves (arxiv:2606.09498): A harness architecture where the control stack autonomously diagnoses its own weaknesses and proposes structural patches, without requiring full re-authoring.
  Relevance to FH: Validates the hub compounding loop / harvest-loop design as an emerging research category; the autonomous self-improvement framing is precisely what FH's 4-axis auto-gate is approximating. Candidate for a Mode D sister-asset cross-audit.

• Claude Code Dynamic Workflows — native parallel orchestration (InfoQ, June 2026): Anthropic shipped Dynamic Workflows, letting Claude dynamically generate orchestration scripts, decompose tasks into subtasks, and run them in parallel — confirmed to be trending on HN (#1 hit).
  Relevance to FH: FH's agent-composer + parallel-dispatch patterns now have a native CC analogue. Warrants a persona-innovator scan to check whether agent-composer SKILL.md can delegate to the native API rather than reinventing it — potential simplification.

• NSA MCP Security Guidance + enterprise gateway architecture (NSA CSI, May 2026; Practical DevSecOps guide): NSA published MCP security guidance formalising tool-allowlisting, centralized gateway proxies (TrueFoundry, IBM ContextForge), and approval gates for destructive tool calls as the enterprise standard. Community consensus is that server-side annotations are unreliable — the guidance aligns with FH's measured finding from 2026-06-11.
  Relevance to FH: Strong external validation for mcp_tool_gating.md's name-keyed ask/allow table. The drift toward centralized gateway proxies suggests a future mcp-circuit-breaker extension or a gateway-pattern note in mcp_tool_gating.md.

• Active Context Compression (arxiv:2601.07190): Autonomous in-session memory management — agents decide when to compact, summarise, or evict context without user prompting, using failure-driven guidelines. The field is converging on "context drift kills agents before context limits do."
  Relevance to FH: /context-doctor is well-placed but currently reactive (user-triggered). An autonomous trigger dimension — proactively proposing compaction when drift signals appear, rather than waiting for /clear keywords — is the gap this paper identifies. Low-friction improvement candidate.

• HN "skills beat prompts" consensus (Developers Digest; codewithandrea): June 2026 HN threads show community convergence that encoding repeatable workflows as skills (reusable, version-controlled, gate-compliant) is the dominant differentiator between 2025 and 2026 AI-agent practices. Anthropic's own docs now use "skills" as the canonical unit of reusable agent work.
  Relevance to FH: FH's skill-first philosophy and 6-axis creation gate are externally validated. The community language now matches FH terminology without internal jargon — FH's natural-language trigger coverage (SKILL.md §Trigger Phrases) likely has lower collision risk than a year ago.

---

📡 HN ~10 items · arxiv 7 papers · WebSearch synthesis · no operator online (scheduled run)

---

Generated by Claude Code

[chrono-meta]

📋 Weekly Routine Summary — 2026-06-22

Digest window read: 2026-06-16 through 2026-06-21 (6 daily digests)

---

Signal selected

O'Reilly AI Agents Stack 2026 (June 21 digest)
https://www.oreilly.com/radar/the-ai-agents-stack-2026-edition/

"Context engineering" is the 2026 practitioner consensus term for the discipline context-doctor covers. None of the three dominant 2026 terms — context engineering, context rot, context collapse — appeared in context-doctor's trigger list. The June 21 digest explicitly flagged: "Update context-doctor triggers to include 'context engineering' as a synonym for external discoverability; the 32K rot threshold is a candidate concrete gate recommendation."

---

What was shipped

Draft PR #112: frontier-auto: context-doctor 2026 trigger vocab + 32K rot threshold
#112

plugins/fh-meta/skills/context-doctor/SKILL.md — 4 lines added:

• §Explicit Triggers: added "context engineering", "context rot", "context collapse"
• §Natural Language Triggers table: new row for the 2026 industry terms
• §Step 3 /clear checklist: □ item for the ~32K token context-rot-onset threshold

Gate results: Axis 1 ✅ PASS · Axis 2+3 ⚠️ below-floor (sonnet tier, weekly-audit re-queue) · Axis 4 ✅ PASS

---

Signals that did NOT meet the bar this week

• harness-doctor L3 state-degradation (Jun 18): already covered in the existing Harness-Defect Taxonomy table — no gap.
• mcp_tool_gating AAIF tier (Jun 21): warning signal itself said "do not prematurely upgrade trust tiers" — deferred.
• agent-composer AdaptOrch topology (Jun 20): valid candidate but requires judgment logic, not a clean annotation-only change; deferred.

---

Autonomous weekly routine · draft PR only · HITL merge required
Generated by Claude Code

---

Generated by Claude Code

[chrono-meta]

Frontier Digest — 2026-06-23

🔑 Engine: WebSearch mode (in-context synthesis) · Sources: HackerNews Algolia + arxiv search · Autonomous collection arm only (no operator intake)

---

Frontier Signals (3–6 bullets, FH-relevant)

• "The Last Harness You'll Ever Build" (arxiv 2604.21003, Sylph.AI, May 2026) — Proposes a three-agent Harness Evolution Loop (Worker → Evaluator → Evolution Agent) that automatically optimizes prompts, tools, and orchestration logic for any new task domain — no expert hand-engineering required.
  Relevance to FH: FH's 4-axis auto-gate (regression → adversarial → forward → record) is structurally isomorphic; this paper's autonomous outer-loop closes the loop FH currently closes manually via harvest-loop + HITL. Prime candidate for a future FH improvement proposal via /field-harvest.

• "Building Effective AI Coding Agents for the Terminal: Scaffolding, Harness, Context Engineering" (arxiv 2603.05344, Mar 2026) — Peer-reviewed treatment of CLAUDE.md, hooks (27 hook events), and compaction stages (5 levels) as the primary harness surface for terminal AI agents.
  Relevance to FH: FH's CLAUDE.md-first architecture and session-wrap close chain are now academically validated; the hook/compaction metrics give concrete benchmarks FH can reference in harness-doctor diagnostics.

• "ContextBudget: Budget-Aware Context Management for Long-Horizon Search Agents" (arxiv 2604.01664) + "The Missing Memory Hierarchy: Demand Paging for LLM Context Windows" (arxiv 2603.09023) — Two converging papers model context as a finite, budgeted resource with demand-paging semantics — evict, page-in, track cost explicitly.
  Relevance to FH: Strengthens the theoretical grounding for FH's token-budget-gate and context-doctor skills; "demand paging" is a useful frame for the LOCA-bench context growth benchmarks — FH could adopt similar budget-tracking in its execution-tier settings.

• "Show HN: Gambit — open-source agent harness for building reliable AI agents" (HN #46641362) — Defines an agent harness as an "operating system for agents" handling tool calling, planning, context window management, and developer orchestration. 400+ points on HN.
  Relevance to FH: Independent convergence on FH's "Control Tower" framing — "harness as OS" is the phrase the wider community is landing on. FH should ensure its public-facing README reflects this shared vocabulary (sister-asset candidate for knowledge/shared/).

• "Configuring Agentic AI Coding Tools: An Exploratory Study" (arxiv 2602.14690) — Empirical study of how teams actually configure CLAUDE.md / AGENTS.md in practice: most teams use 3–7 rule files, overrides dominate, and configuration drift is the fix(meta-validation): Wave 3 — table count, CI agent detect, path neutralize + external network artifacts #1 reported pain point.
  Relevance to FH: FH's harness-doctor L1–L4 diagnosis and drift-prevention principles address exactly this pain class. Configuration drift being the top pain point validates FH's Harness Drift Prevention Principles section — and suggests a more prominent harness-doctor cadence proposal at onboarding.

• HN: "The current hype around autonomous agents, and what actually works in production" (#44623207) — Top-voted community thread; recurring theme: agent reliability degrades sharply without explicit failure modes, human-in-the-loop gates, and verification loops. "Brute-forcing reliability with bigger models and longer prompts" (Statewright / HN #48108778) is the dominant but fragile pattern in the wild.
  Relevance to FH: FH's HITL gates (feedback_no_personal_commit_to_shared_repo, Pre-Publish Surface Gate, Destructive-Op Gate) and steel-quench's "simpler alternative?" first-wave attack are precisely the defensive patterns production teams are rediscovering the hard way. Good calibration signal — FH is ahead of the median production pattern.

---

📡 Collection stats: HN ~10 items scanned · arxiv ~9 papers scanned · TLDR/Batch: not fetched (WebSearch mode)
🔗 Autonomous daily run — no operator intake (Step 0.5 skipped per task spec)

---

Generated by Claude Code

[chrono-meta]

Frontier Digest — 2026-06-24

🔑 Engine: WebSearch mode (autonomous collection arm — HackerNews + arxiv)

---

• "Agent = Model + Harness" reaches Gartner-tracked industry consensus. Model capability has converged at the frontier; differentiation has shifted entirely to the harness layer. Gartner projects 40% of enterprise apps will include task-specific AI agents by EOY 2026 — the harness category is now classified as infrastructure, not optional tooling.
  Source: Agent = Model + Harness — Cobus Greyling, Jun 2026
  Relevance to FH: Validates FH's Core Axis premise and dual-layer architecture. FH's "simpler over time" discipline is the correct posture for infrastructure-grade harnesses.

• DACS: Dynamic Attentional Context Scoping (arxiv 2604.07911, Apr 2026). Formal multi-agent context isolation mechanism: orchestrator holds lightweight per-agent registry summaries (≤200 tokens) in REGISTRY mode; on an agent's STEERING REQUEST, the orchestrator injects that agent's full context and compresses all others. 90–98% steering accuracy vs 21–60% for flat-context baseline; wrong-agent contamination drops from 28–57% to 0–14% across 200 trials.
  Source: arxiv 2604.07911
  Relevance to FH: Directly validates FH's Context Card pattern and the [Session Context Card] dispatch format. REGISTRY/Focus is a measured, citable version of the same isolation idea. Candidate for FH's knowledge/shared/ as an empirical anchor for the dispatch protocol.

• Agentjacking: MCP-based attack class documented (The New Stack, Jun 17 2026). Tenet Security disclosed an attack where forged Sentry MCP events trick Claude Code and Cursor into executing attacker-controlled code — a concrete exploit of prompt injection through a mounted MCP server's tool output.
  Source: A public Sentry key is all it takes to hijack Claude Code — The New Stack
  Relevance to FH: Confirms the threat model behind templates/.claude/rules/mcp_tool_gating.md — the name-keyed ask/allow table and "never trust server annotations" rule were correctly anticipated. FH's Pre-Publish Gate and MCP gating are now empirically justified, not just precautionary.

• CLAUDE.md / AGENTS.md converging as a cross-tool convention (HN, Jun 2026). CLAUDE.md (Claude Code-specific) and AGENTS.md (cross-tool) are becoming the standard mechanism for encoding project rules and operational expectations so any agent can reuse them without restating context per session. "Claude Code started as a terminal assistant and now runs as a layered agentic system separating memory, hooks, skills, subagents, plugins, and MCP into distinct layers."
  Source: What HN Gets Right About AI Coding Agents in 2026
  Relevance to FH: FH's templates/CLAUDE.md propagation path is ahead of this curve. The FH AGENTS.md file is well-positioned as the Codex-compatible entry point. Signal to watch: if AGENTS.md becomes a formal cross-tool spec, FH's templates/ should include it as a first-class install item.

• ContextBudget: budget-aware context management for long-horizon agents (arxiv 2604.01664). Introduces formal budget constraints on context selection for search agents — models that explicitly track "context budget remaining" and gate on it outperform those that truncate at overflow.
  Source: arxiv 2604.01664
  Relevance to FH: Aligns with FH's /context-doctor + /token-budget-gate skill pair. The formal budget-constraint framing could strengthen token-budget-gate's gate criteria — current FH implementation is prose-driven; a budget-ceiling anchor would harden it.

• Natural-Language Agent Harnesses: formal academic treatment (arxiv 2603.25723). A paper explicitly titled "Natural-Language Agent Harnesses" signals that FH's methodology territory is now arxiv-tracked. Formal characterization of harnesses as distinct from raw agents.
  Source: arxiv 2603.25723
  Relevance to FH: Sister-asset candidate for knowledge/shared/harness-core/. Activates the Sister Asset Protocol — worth a cross-audit session to confirm overlap / import candidates before the next harvest-loop run.

---

📡 Collection stats: HN (WebSearch) · arxiv (WebSearch) · synthesis: in-context (WebSearch mode, no API key)
Autonomous daily scan — no operator online. Chaining deferred (Step 4 skipped per task constraints).

---

Generated by Claude Code

[chrono-meta]

Frontier Digest — 2026-06-25

🔑 Engine: WebSearch mode (no ANTHROPIC_API_KEY — synthesized in-context)
📡 Sources: HackerNews · arxiv · secondary references

---

Signals (3–6 bullets, FH-relevant)

• "Code as Agent Harness" survey lands (arXiv:2605.18747, May 2026) — 40+ authors (UIUC, Meta, Stanford) frame code as the operational substrate for agent reasoning, acting, and verification. Three-layer model: harness interface → harness mechanisms (planning/memory/tool use) → execution-based feedback control. Directly validates FH's 6-axis framework and the "harness is the system" thesis; the paper's open-challenge list (evaluation beyond final-task success, regression-free harness improvement, consistent shared state across agents) maps directly onto FH Axis-5 check-classes and the 4-axis pre-commit gate.
  — Source: https://arxiv.org/abs/2605.18747

• "Externalization in LLM Agents" unified review (arXiv:2604.08224, Apr 2026) — Shanghai Jiao Tong / CMU / OPPO survey treats memory, skills, protocols, and harness engineering as a single externalization framework: "harness engineering serves as the unification layer that coordinates them into governed execution." Strongest academic framing yet of what FH's methodology layer calls the 6-axis. Candidate for sister-asset cross-audit (tracks/_audit/); the framing of skills-as-externalized-procedural-expertise aligns with FH's SKILL.md gate and the no-reinvention rule.
  — Source: https://arxiv.org/abs/2604.08224

• HN: "Improving 15 LLMs at Coding in One Afternoon. Only the Harness Changed" — HN thread surfacing empirical evidence that harness changes move the needle more than model upgrades. Reinforces FH's founding axiom and could be cited in the hub's README or onboarding to explain why harness investment compounds. Useful input for the next frontier-digest --chain → /field-harvest cycle.
  — Source: https://news.ycombinator.com/item?id=46988596

• HN: "Karpathy-style LLM wiki your agents maintain (Markdown + Git)" — LLM-native knowledge substrates where agents read and write so context compounds across sessions rather than being re-pasted. Closest external cousin to FH's knowledge/ compounding loop and MEMORY.md design. Lightweight sister-asset check recommended: C-tier if territory is already covered by hub_compounding_loop.md; A-tier if the Git-native write-back mechanism is novel.
  — Source: https://news.ycombinator.com/item?id=47899844

• Claude Code deterministic multi-agent orchestration pattern maturing — Community documentation (alexop.dev, codewithandrea.com) is now treating Claude Code Workflows (parallel sub-agents, deterministic fan-out) as a stable pattern. The most advanced users build harnesses around it: CLAUDE.md, hooks, MCP integrations, review pipelines, CI workflows — exactly the FH Full-Harness Mode asset set. Signal: the gap between "Claude Code power user" and "FH-structured project" is narrowing; FH's differentiation is now governance depth (4-axis gate, mechanical anchors) rather than workflow patterns alone.
  — Source: https://alexop.dev/posts/claude-code-workflows-deterministic-orchestration/

• "Synthesizing Multi-Agent Harnesses for Vulnerability Discovery" (arXiv:2604.20801) — Optimizer produces a single task-agnostic harness program (same AgentFlow graph, prompts, tool bindings) deployed identically across tasks; a diagnoser observes per-task outcomes to localize harness weaknesses. Pattern directly relevant to FH's regression_guard (Axis-1) and the concept of a harness that self-diagnoses and self-improves — watch as a potential /harness-doctor upgrade signal.
  — Source: https://arxiv.org/abs/2604.20801

---

Collection stats: HN 4 items · arxiv 6 items · secondary 3 items
Chaining: 3+ Immediate Application Candidates found — recommend next /field-harvest or fh_signal save on the externalization taxonomy and Code-as-Harness open challenges.

---

Generated by Claude Code

[chrono-meta]

Frontier Digest — 2026-06-26

Autonomous daily collection · WebSearch mode · HackerNews + arxiv · synthesized in-context

---

Signals (3–6 bullets, FH-relevant)

• CI gates as a harness-first concept — "LLM Readiness Harness: Evaluation, Observability, and CI Gates for LLM/RAG Applications" (arXiv 2603.27355) formalizes scenario-weighted readiness scores and CI quality gates that block risky prompt variants before release.
  Relevance to FH: Validates and mirrors FH's 4-axis pre-commit gate (Axis 1 regression + Axis 5 check-class); the "block on fail, not just report" posture is the same design principle FH's hook enforces.

• Multi-agent transactive memory as a compounding loop — "Multi-Agent Transactive Memory" (arXiv 2606.19911, CMU/UC Berkeley, June 2026) proposes population-level storage and retrieval of agent-generated trajectories so consumer agents reuse producer-agent knowledge.
  Relevance to FH: Direct structural match to FH's cross-project skill bus + harvest-loop compounding loop; the paper's producer/consumer framing is a formalised version of FH's field → hub sync circuit.

• Runtime harness adaptation without touching the model — "Adapting the Interface, Not the Model: Runtime Harness Adaptation for Deterministic LLM Agents" (arXiv 2605.22166) introduces Life-Harness, converting recurring interaction failures into reusable harness interventions across 18 model backbones.
  Relevance to FH: Empirical evidence for FH's own layered design (hook-mechanical enforcement vs prose-advisory split); the "fix the harness, not the model" framing is the engineering rationale behind FH's pre-commit gate being hook-enforced rather than prompt-advisory.

• Claude Code multi-agent swarm ecosystem maturing rapidly — HN cluster (Feb–June 2026): Claude Code Swarms · ClaudeSwarm OSS · 20+ coordinating agents · Lessons from swarms at scale. Multiple open-source Claude Code orchestrators emerged and Anthropic shipped native Agent Teams.
  Relevance to FH: FH's parallel agent dispatch pattern is on the correct trajectory; the "lessons at scale" thread likely surfaces failure modes (stalling, loops, non-compiling output) that directly inform FH's Context Card and self-contained sub-agent prompt rules.

• Skill evaluation emerging as its own discipline — "Agent Skill Evaluation and Evolution: Frameworks and Benchmarks" (arXiv 2606.11435, June 2026) surveys evaluation-driven skill evolution across four paradigms (execution feedback, trajectory distillation, compression, RL) and flags structural gaps in benchmark coverage.
  Relevance to FH: FH's New Skill Creation Pre-Commit Gate (check-class declared · Done-When defined · natural language triggers ≥ 3) is a lightweight instantiation of the evaluation-driven skill evolution approach; the paper's "benchmark coverage gaps" angle maps to FH's harness-doctor L1–L4 diagnosis.

---

Engine: WebSearch (in-context synthesis) · Sources: arxiv.org, news.ycombinator.com · Collection run: 2026-06-26

---

Generated by Claude Code