Prompt engineering is a suggestion. Harness Kit is a guarantee.
Quick Start • The Problem • Architecture • Features • Benchmarks
Every team building production AI agents rediscovers the exact same painful lessons independently. You build an agent. It works flawlessly in a 5-minute Slack demo. You put it in production, and:
- ❌ Token Death Spirals: The agent tries a
bashcommand. It fails. It tries the exact same command 14 times in a row, burning $50 in tokens over the weekend. - ❌ Context Amnesia: The context window fills up, compaction triggers, and the agent completely forgets what files it already edited, abandoning the feature halfway.
- ❌ Rule Ignoring: You write an exhaustive
AGENTS.mdtelling it to "never use raw HTML." It creatively ignores you because LLMs treat prompts as suggestions, not constraints. - ❌ Over-tooling: You give the agent 15 tools to be helpful. It gets confused, hallucinates arguments, and crashes.
The hard truth: You cannot fix non-deterministic models with more prompt engineering. You need a mechanical operating system.
Harness Kit is the missing infrastructure layer between "I have an LLM" and "I have a reliable agent."
It treats the LLM as a raw, unreliable CPU and acts as the strict Operating System—sandboxing its tools, mechanically enforcing its budgets, and persisting its memory outside the context window.
Built on production lessons from Stripe's Minions, Anthropic, Vercel, and recent arXiv research, Harness Kit is a high-performance Rust core wrapped in a seamless DX.
| Approach | Success Rate | Tokens Burned | Doom Loops |
|---|---|---|---|
| Raw Agent (GPT-4o / Opus) | 58% | 140k | ~15% |
| With Harness Kit | 96% | 61k | 0% |
Run harness score --suite standard to reproduce these numbers locally.
You don't need to rewrite your agent. Drop Harness Kit into your existing repo.
# 1. Initialize Harness Kit in your repo
npx harness-kit init --analyze
# 2. Run a fully autonomous, reliable agent session
npx harness-kit run "Build a complete Stripe checkout flow and fix any test failures"Watch it work. It initializes, writes a machine-readable feature manifest, runs coding sessions, mechanically halts on errors, commits progress to git, and tests each feature. It just doesn't fail.
Ever wonder why your agent succeeded (or failed)? Run:
npx harness-kit replayGenerates a shareable, visual timeline video of exactly what your agent did: which tools it called, where it got stuck, how it recovered, and the git diff at every step.
Harness Kit isn't a wrapper. It is 5 distinct, highly-opinionated subsystems.
- What it prevents: Agents getting lost on complex tasks.
- How: Implements hybrid workflows (Stripe's Minions). Scaffolding (cloning, running linters, opening PRs) is done via deterministic, token-free code. Only the genuine creative work (writing code, fixing tests) is handed to the LLM.
- What it prevents: $500 token-burning death spirals.
- How: Intra-session guardrails physically count identical tool failures. If an agent tries the same broken
bashcommand 3 times, it triggers a hard interrupt and injects an error context. Limits CI test-fix loops to 2 max.
- What it prevents: The 3% performance drop caused by bloated
AGENTS.mdfiles (arXiv: 2602.11988). - How: Scans your project instruction files and violently strips out redundant LLM-generated prose. It enforces that context files only contain concrete constraints (e.g., "Use
npm run test, notjest"). UsesScopedContextInjectorto only feed the agent rules for the exact directory it is currently editing.
- What it prevents: Agents wiping your hard drive or hallucinating outdated APIs.
- How: Uses native Linux namespaces (
--net,--ipc,--pid) to completely isolatebashexecutions. Implements a strictToolBudget—hard-capping how many tools the agent can use per session. Includes nativeLSPsupport for semantic code search, reducing grep-based hallucinations. - Bonus: Native integration with Andrew Ng's
context-hubto fetch external API docs dynamically, preventing API hallucination.
- What it prevents: Agents hallucinating task completion or forgetting state.
- How: Splits memory into Episodic (cross-session persistent logs) and Working (current context window). Uses a strictly-typed, write-protected
FeatureManifest.json—the agent can only flip flags totrue, it cannot creatively rewrite your spec. Universal undo viaGitCheckpointeron every failed loop.
Harness Kit is Model & Runtime Agnostic. It does not replace your AI SDK; it wraps it.
import { Harness } from 'harness-kit';
import { generateText } from 'ai'; // Vercel AI SDK, LangGraph, etc.
const harness = new Harness({
model: 'claude-3-5-sonnet',
budget: { maxTools: 4, maxCiRetries: 2 },
tools: ['fs', 'bash', 'lsp', 'context-hub']
});
await harness.run("Migrate the database to PostgreSQL");🤝 Context-Hub Integration: Works perfectly out-of-the-box withcontext-hub to ensure your agents never write code against deprecated, hallucinated API endpoints.
- Enforcement over Suggestion: If a rule isn't mechanically enforced in the runtime, it doesn't exist.
- File System is Truth: No external databases. Your state, manifests, and progress logs live in your repo and are git-versioned.
- Fewer Tools Win: Vercel proved 3 tools beat 15. The library actively makes it harder to add unnecessary tools.
- Rippable by Design: Plugs into LangChain, LangGraph, or native SDKs. As frontier models get smarter, eject any module you no longer need.
We are actively porting advanced components into the native Rust core (crates/runtime). If you want to help build the future of reliable agentic infrastructure, see our CONTRIBUTING.md.
If Harness Kit saved you from a token death spiral, consider giving us a ⭐.