Who it is forA researcher studying agent self-improvement who needs concrete examples of edits, reverts, and lessons.
The problemClaims about agent learning are hard to check without seeing each change the agent made and why.
The outcomeA list of strategy edits with the retro reasoning for keeping, sharpening, or reverting each one.
What you need first
- Python 3 (CI runs on Python 3.12); the engine uses the standard library only
- git and a clone of the repository
- A Claude Code account and the `claude` command on your PATH, for running cycles
- A fork with write access to origin, or no origin remote for local-only runs
- Pearl Connect installed from pearl.you/connect and a funded Safe, only for real-money mode
Steps
git clone https://github.com/bennyjo/phil.gitcd philgit log --oneline -- strategy/to list strategy commits. (inferred, check the README)- Open
journal/retros/and find the daily deep-retro entries that audit those edits. - Compare each kept or reverted edit with the settled bets it cites.
Prompt for your agent
Clone https://github.com/bennyjo/phil, list commits that touch strategy/, and match each to a retro in journal/retros/. Output a table of edit, decision (keep, sharpen, or revert), and the evidence cited. Verify that every row cites a retro file that exists.
Caveat. The README says the git log is the record of lessons, but a single edit's effect on results is hard to isolate from the other changes made in the same period.
Use this in your AI agent
Ask your agent
Using the Deduped MCP, get the setup steps for bennyjo/phil and "Trace which strategy edits Phil kept or reverted"
Add Deduped to Claude Code once
claude mcp add --transport http deduped https://dedupedai.com/mcp
Every Monday: the 5 hottest repos and exactly how to use them. Free, one email a week, unsubscribe anytime.
Get them freeShare: X LinkedIn Hacker News Email Copy link
Third-party repos we don't own. AI-written use cases can be wrong. Disclaimer