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.
Evolution Arena — 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: Arcade Run (Galaga archive) — phase 1 shipped and playtested before the pivot.
Current work, completed projects, and prototypes. Filter by category to find what you're looking for.
Each entry below is a real handoff chain from the agent-brain system, curated down to the ones worth showing. Root nodes are dispatches to Claude Code, from Grok or from a Hermes cron tickler. Delegations are Claude Code→Codex sub-tasks. Click any row to see details.
Data source: agent-brain/watcher/chain-registry.json (private repo) — curated static snapshot, regenerated on deploy. Roots flagged public in the registry are the ones shown.
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?
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.
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.
August 9, 2026: the fully unattended version ran. Grok dispatched Claude Code, Claude Code closed its step and opened the next handoff, headless Grok wrote the receipt, Hermes told me on Telegram. Nobody watched it happen. Between July and August the same machinery ran nine weekly skills reviews and two monthly review days from cron ticklers.
September 2026: I retired Hermes. The watcher and its crons went quiet with it. The pattern survives (handoff files, registry, completion files), but dispatch now starts from a Claude Code or Grok session instead of a cron tick. Which jobs get re-hosted, and where, is the current open question.
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.
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.
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.
codex exec 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.
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.
Hermes could send a Telegram message about what an agent did, but the agent that produced the message didn't have Hermes's delivery context and vice versa. The coordination layer (handoffs, registry) is the shared ground truth.
When Hermes came down, nine documented crons went with it and nothing complained. The registry still showed 22 chains as "dispatched" that had closed weeks earlier. Creation is not verification. Now every job needs an observed live run, and the site check fails if the numbers on this page drift from the data.
Things I haven't figured out yet. These are genuine open questions, not rhetorical ones.
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?
untestedIf 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?
exploringThe 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?
openHermes owned the Telegram capture, the crons, and the watcher. With it gone, which layer owns scheduling and capture: Grok Bot routines, Claude Code scheduled runs, or plain systemd timers? The second-use rule says wait for a real need before rebuilding it.
open