-
Notifications
You must be signed in to change notification settings - Fork 0
Expand file tree
/
Copy pathexperiments.html
More file actions
326 lines (296 loc) · 15.4 KB
/
Copy pathexperiments.html
File metadata and controls
326 lines (296 loc) · 15.4 KB
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
<!DOCTYPE html>
<html lang="en">
<head>
<meta charset="UTF-8">
<meta name="viewport" content="width=device-width, initial-scale=1.0">
<meta name="referrer" content="strict-origin-when-cross-origin">
<meta http-equiv="Content-Security-Policy" content="default-src 'self'; script-src 'self'; style-src 'self' 'unsafe-inline' https://fonts.googleapis.com; font-src https://fonts.gstatic.com; img-src 'self' data:; object-src 'none'; base-uri 'self'; frame-ancestors 'none'; form-action 'self'">
<title>Experiments — griffithee</title>
<meta name="description" content="AI agent delegation chains, autonomy experiments, and what I'm learning.">
<link rel="preconnect" href="https://fonts.googleapis.com">
<link rel="preconnect" href="https://fonts.gstatic.com" crossorigin>
<link href="https://fonts.googleapis.com/css2?family=Inter:wght@400;500;600;700&family=JetBrains+Mono:wght@400;600&display=swap" rel="stylesheet">
<link rel="stylesheet" href="css/style.css">
</head>
<body>
<nav>
<div class="nav-inner">
<a href="index.html" class="nav-logo">griffith<span>ee</span></a>
<ul class="nav-links">
<li><a href="index.html">Home</a></li>
<li><a href="projects.html">Projects</a></li>
<li><a href="experiments.html">Experiments</a></li>
<li><a href="game.html">Game</a></li>
<li><a href="about.html">About</a></li>
</ul>
</div>
</nav>
<main>
<div class="container">
<header class="page-header">
<div class="page-header-eyebrow">Experiments & Learning</div>
<h1 class="page-header-title">What I'm figuring out</h1>
<p class="page-header-desc">
Working notes on the autonomy experiments. What's working, what isn't,
and what surprised me. The chain visualizer below shows real delegation
data from the agent system.
</p>
</header>
<div class="callout" style="margin-bottom:var(--space-8);">
<div class="callout-label">Current live experiment</div>
<p style="margin-bottom:0;">
<a href="game.html">Evolution Arena</a> — 40 agents with tiny neural networks
learn to find food via a genetic algorithm, entirely in your browser. No API calls,
no pre-trained model. The question: how fast does selection pressure produce
genuinely useful behavior from random noise?
Prior experiment: <a href="galaga.html">Arcade Run (Galaga archive)</a> — phase 1 shipped and playtested before the pivot.
</p>
</div>
<!-- Projects at a Glance -->
<section class="section" id="projects-browser">
<div class="section-header">
<div class="section-label">Overview · All work</div>
<h2 class="section-title">Projects & Experiments at a Glance</h2>
<p class="section-desc">
Current work, completed projects, and prototypes. Filter by category to find what you're looking for.
</p>
</div>
<div id="projects-browser-container">
<div style="color:var(--text-muted);font-family:var(--font-mono);font-size:0.8rem;">Loading…</div>
</div>
</section>
<hr class="divider">
<!-- Chain Visualizer -->
<section class="section" id="chain-visualizer">
<div class="section-header">
<div class="section-label">Interactive · Live data</div>
<h2 class="section-title">Agent Delegation Chain Visualizer</h2>
<p class="section-desc">
Each entry below is a real handoff chain from the agent-brain system.
Root nodes are Grok→Claude Code dispatches. Delegations are Claude Code→Codex sub-tasks.
Click any row to see details.
</p>
</div>
<div class="visualizer-wrap">
<div class="visualizer-toolbar">
<span class="visualizer-title">watcher/chain-registry.json · snapshot loading…</span>
<div class="filter-btns">
<button class="filter-btn active" data-filter="all">all</button>
<button class="filter-btn" data-filter="active">active</button>
<button class="filter-btn" data-filter="closed">closed</button>
</div>
</div>
<div class="visualizer-body">
<div id="chain-list-container">
<div style="color:var(--text-muted);font-family:var(--font-mono);font-size:0.8rem;">
Loading chain data…
</div>
</div>
<div id="chain-detail-panel" class="chain-detail-panel">
<!-- populated by visualizer.js -->
</div>
</div>
</div>
<div style="margin-top:var(--space-4);">
<p style="font-size:0.8rem;color:var(--text-muted);font-family:var(--font-mono);">
Data source: agent-brain/watcher/chain-registry.json — static snapshot, updated on deploy.
<a href="https://github.com/griffithee/agent-brain" target="_blank" rel="noopener">View the source system →</a>
</p>
</div>
</section>
<hr class="divider">
<!-- Autonomy experiments narrative -->
<section class="section" id="autonomy">
<div class="section-header">
<div class="section-label">Experiment 1</div>
<h2 class="section-title">Autonomous multi-agent delegation</h2>
</div>
<div class="highlight-row">
<div>
<p>
The core question: can I hand a goal to Grok, have Grok hand it to Claude Code,
Claude Code hand sub-tasks to Codex, and have the whole loop close — including
the git push and notification — without me stitching anything together?
</p>
<p>
As of May 2026: yes, for the single-chain case. The watcher script monitors
the handoffs directory, dispatches agents when files appear, and auto-closes
the loop when completion handoffs are detected.
</p>
<p>
What surprised me: the blockers were all prompt and harness issues, not
capability gaps. Claude Code was capable of closing loops the whole time —
but early prompt instructions told it to hand off to Codex instead. Once
that was fixed, it worked.
</p>
</div>
<div>
<div class="timeline">
<div class="timeline-item done">
<div class="timeline-date">Early May 2026</div>
<div class="timeline-title">First handoff experiment</div>
<div class="timeline-desc">Manual Grok→Claude Code handoff. Claude Code closes with a commit and completion file. Manual stitching.</div>
</div>
<div class="timeline-item done">
<div class="timeline-date">Mid-May 2026</div>
<div class="timeline-title">Watcher automates dispatch</div>
<div class="timeline-desc">handoff-watcher.sh dispatches Claude Code automatically when files appear in handoffs/.</div>
</div>
<div class="timeline-item done">
<div class="timeline-date">May 30, 2026</div>
<div class="timeline-title">Full chain closes without human</div>
<div class="timeline-desc">Grok→Claude Code→Codex→Claude Code→Grok completes end-to-end. Chain registry + Strategy R.</div>
</div>
<div class="timeline-item done">
<div class="timeline-date">May 31, 2026</div>
<div class="timeline-title">Session limit detection</div>
<div class="timeline-desc">Watcher detects usage limit hits, sends Telegram alert, auto-retries next cycle.</div>
</div>
<div class="timeline-item done">
<div class="timeline-date">May 31, 2026</div>
<div class="timeline-title">Website as live test</div>
<div class="timeline-desc">Claude Code builds this site as orchestrator. Phase 1 (foundation), Phase 2 (visual polish), Phase 3 (living artifact) all closed autonomously.</div>
</div>
<div class="timeline-item done">
<div class="timeline-date">May 31, 2026</div>
<div class="timeline-title">Living artifact pattern</div>
<div class="timeline-desc">Phase 3: chains.json refresh script, filterable projects browser, update ritual. Site maintainable in <20 min without re-engaging the full agent system.</div>
</div>
<div class="timeline-item active">
<div class="timeline-date">Jun 10, 2026 — now</div>
<div class="timeline-title">Neuroevolution arena</div>
<div class="timeline-desc">Galaga replaced with a zero-cost AI experiment: 40 neural-net agents evolved by genetic algorithm, competition mechanics, interactive food placement.</div>
</div>
<div class="timeline-item done">
<div class="timeline-date">Jun 10, 2026</div>
<div class="timeline-title">Vault maintenance automation</div>
<div class="timeline-desc">Claude Code builds a weekly automated health check for the 1,700+ file knowledge vault. Four deterministic checks: broken wikilinks, orphaned notes, frontmatter drift, book summary gaps. Weekly Hermes cron, zero manual work.</div>
</div>
<div class="timeline-item active">
<div class="timeline-date">Jun 2026 — now</div>
<div class="timeline-title">SaveWisdom archive at 103</div>
<div class="timeline-desc">103 reflection questions answered and archived across two stores (Obsidian vault + agent-brain). System maintained through vacation via same-day resend protection. Dual-persistence pipeline solid.</div>
</div>
<div class="timeline-item active">
<div class="timeline-date">Jun 28, 2026 — now</div>
<div class="timeline-title">fitness-tracker live</div>
<div class="timeline-desc">Telegram food, workout, and weight logging is deployed with noon tickles and nightly reports. InBody scan metadata import remains the next cleanup item.</div>
</div>
</div>
</div>
</div>
</section>
<hr class="divider">
<!-- Learnings -->
<section class="section" id="learnings">
<div class="section-header">
<div class="section-label">Key findings</div>
<h2 class="section-title">What I've learned</h2>
</div>
<div class="card-grid">
<article class="card">
<div class="card-title" style="margin-bottom:var(--space-3);">Capability gaps are usually prompt gaps</div>
<p class="card-desc">
Several times I assumed an agent couldn't do something and built a workaround —
only to find the agent could do it fine, but the prompt was instructing it not to.
Debug the instructions before debugging the model.
</p>
</article>
<article class="card">
<div class="card-title" style="margin-bottom:var(--space-3);">Live tests find what static review misses</div>
<p class="card-desc">
Code that passes syntax checks and static review can still fail in the
watcher's cron environment. PATH differences, prompt/detector mismatches,
timing issues — these only appear when the system actually runs.
</p>
</article>
<article class="card">
<div class="card-title" style="margin-bottom:var(--space-3);">Declare the role in the prompt</div>
<p class="card-desc">
Agents need to know their role in the current chain. "You are the orchestrator"
vs "you are the sub-task implementer" changes behavior significantly. Same model,
different stance.
</p>
</article>
<article class="card">
<div class="card-title" style="margin-bottom:var(--space-3);">Codex non-interactive is not the closer</div>
<p class="card-desc">
<code>codex exec</code> in non-interactive mode reads the prompt and exits
without creating handoffs or pushing. Useful for bounded code-gen. Not useful
for loop-closing, review, or orchestration.
</p>
</article>
<article class="card">
<div class="card-title" style="margin-bottom:var(--space-3);">Registry beats filename parsing</div>
<p class="card-desc">
Early chain tracking parsed filenames to resolve parent-child relationships.
Fragile. The chain registry (Strategy R) looks up delegation filenames directly —
works even after archiving, collision-safe for parallel chains.
</p>
</article>
<article class="card">
<div class="card-title" style="margin-bottom:var(--space-3);">Agent output is not the same as agent context</div>
<p class="card-desc">
Hermes can send a Telegram message about what an agent did, but the agent
that produced the message doesn't have Hermes's delivery context and vice versa.
The coordination layer (handoffs, registry) is the shared ground truth.
</p>
</article>
</div>
</section>
<hr class="divider">
<!-- What's next -->
<section class="section" id="next">
<div class="section-header">
<div class="section-label">What's next</div>
<h2 class="section-title">Open questions</h2>
<p class="section-desc">
Things I haven't figured out yet. These are genuine open questions, not rhetorical ones.
</p>
</div>
<div class="card-grid">
<article class="card" style="border-style:dashed;">
<div class="card-title" style="margin-bottom:var(--space-3);">Parallel chains</div>
<p class="card-desc">
Strategy R is designed to be collision-safe for concurrent chains, but I haven't
run a real parallel load test. Does it actually work when two Grok→Claude dispatches
are active simultaneously?
</p>
<span class="tag tag-orange">untested</span>
</article>
<article class="card" style="border-style:dashed;">
<div class="card-title" style="margin-bottom:var(--space-3);">Codex interactive mode</div>
<p class="card-desc">
If Codex interactive mode were reliably triggerable from cron, the Codex role
could expand significantly. Currently it's only useful for bounded, self-contained
code-gen. What would the expanded model look like?
</p>
<span class="tag tag-orange">exploring</span>
</article>
<article class="card" style="border-style:dashed;">
<div class="card-title" style="margin-bottom:var(--space-3);">User intent vs agent output</div>
<p class="card-desc">
The agents are good at executing handoffs. They're less good at knowing when
a handoff's intent has shifted since it was written. How do you route that
kind of ambiguity back to the user without blocking the whole chain?
</p>
<span class="tag tag-orange">open</span>
</article>
</div>
</section>
</div>
</main>
<footer>
<div class="footer-inner">
<span class="footer-text">griffithee · built with Claude Code</span>
<div class="footer-links">
<a href="https://github.com/griffithee" target="_blank" rel="noopener">GitHub</a>
<a href="about.html">About</a>
</div>
</div>
</footer>
<script src="js/nav.js"></script>
<script src="js/projects-browser.js"></script>
<script src="js/visualizer.js"></script>
</body>
</html>