SutroYaro
AGENTS.md16 starsChanged 6 months agoUnlicenseView on GitHub
This project uses AI agents (Claude Code, Codex CLI, Gemini CLI, Replit, others) for research and accepts contributions from both humans and agents. If you are the only agent working on this project, read these in order: Note: CLAUDE.md is the canonical technical source regardless of which model/tool you use (Claude, Gemini, Kimi, etc.). Phase 1 (16 experiments): Single Claude Code sessions running experiments sequentially, each following the template in src/sparse_parity/experiments/_template.py. Phase 2 (17 experiments): 17 independent Claude Code agents…
Where files are headingNew AGENTS.md files have more rules about keeping secrets out of the code, and fewer about code style.
Of the AGENTS.md files we read, the 33,315 first seen in the last six months, against the older ones: the kind of rule that gained most share, and the kind that lost most. About files like this in general, not this one, and an estimate, since a kind is a pattern in the words.
In this repository9 other files it gives its agents.
- .codex/AGENTS.mdAGENTS.md
- CLAUDE.mdCLAUDE.md
- anti-slop-guideSkill
- info-defragSkill
- prepare-meetingSkill
- run-experimentSkill
- sutro-contextSkill
- sutro-syncSkill
- weekly-catchupSkill
The file
We split this AGENTS.md into the rules it gives an agent
12 rules. About 0 are common, 8 are distinctive, and 4 are very specific to this project.
≈0
Common · estimate
Many files like this have a rule of this kind.
≈8
Distinctive · estimate
This author's own choices. Worth a look.
4
Very specific
Tied to their project's names, files or tools. Probably skip.
- Common
- Distinctive
- Very specific
- Distinctive: line 7
If you are the only agent working on this project, read these in order:
- Distinctive: line 28
When reviewing a contributed experiment:
- Very specific: line 30
Check that the experiment ran (results.json should match the findings doc)
- Distinctive: line 31
Check DISCOVERIES.md for prior work on the same question
- Very specific: line 32
Verify the contributor didn't modify measurement code (tracker.py, cache_tracker.py, data.py, config.py)
- Distinctive: line 33
If the experiment answers an open question, check that DISCOVERIES.md is updated
- Very specific: line 38
Follow the template in src/sparse_parity/experiments/_template.py
- Very specific: line 40
Write findings to findings/exp_{name}.md using findings/_template.md
- Distinctive: line 41
Never modify measurement code (LAB.md rule #9)
- Distinctive: line 62
DISCOVERIES.md as shared memory: No agent repeated known-bad configurations (LR=0.5, GrokFast).
- Distinctive: line 64
Strict output format: All 33 experiments are directly comparable because every agent used the same structure.
- Distinctive: line 69
Pyright import warnings: Every agent triggered false "Import could not be resolved" errors (PYTHONPATH is set at runtime).
≈ “Common” is an estimate: at least 1 in 5 of the 42,283 AGENTS.md files we read have a rule of the same kind, not the same words. “Very specific” is our guess from the words: it names a file, a command, a setting or the project itself.
Files
AGENTS.md
This project uses AI agents (Claude Code, Codex CLI, Gemini CLI, Replit, others) for research and accepts contributions from both humans and agents.
For single agents (getting started)
If you are the only agent working on this project, read these in order:
- CLAUDE.md - Deep technical context: project goals, metrics (ARD/DMC), best methods, 36 experiments, current state
- DISCOVERIES.md - Knowledge base: proven facts, failed approaches, open questions (Q7, Q11-Q13)
- LAB.md - Experiment protocol: templates, lifecycle, baselines, rules
- AGENT.md (optional) - Only if running autonomous overnight loops
Note: CLAUDE.md is the canonical technical source regardless of which model/tool you use (Claude, Gemini, Kimi, etc.).
How AI agents were used
Phase 1 (16 experiments): Single Claude Code sessions running experiments sequentially, each following the template in src/sparse_parity/experiments/_template.py.
Phase 2 (17 experiments): 17 independent Claude Code agents dispatched in parallel, each implementing a different algorithmic approach to sparse parity. Each agent received the approach description, shared module APIs, three test configs, and a findings template. All 17 completed successfully, producing code, results, and findings.
Survey: A single agent wrote the Practitioner's Field Guide (docs/research/survey.md) synthesizing all 33 experiments, with spec compliance and code quality review passes.
Meeting #8 onward: Multiple group members run their own agent harnesses (Germain's Replit Research OS, Michael's Claude approach, Yaroslav's Gemini). Results flow in via PRs and the contributions/ directory.
For agents reviewing PRs
When reviewing a contributed experiment:
- Check that the experiment ran (results.json should match the findings doc)
- Check DISCOVERIES.md for prior work on the same question
- Verify the contributor didn't modify measurement code (tracker.py, cache_tracker.py, data.py, config.py)
- If the experiment answers an open question, check that DISCOVERIES.md is updated
For agents running experiments
- Read DISCOVERIES.md first
- Follow the template in
src/sparse_parity/experiments/_template.py - Report both ARD and DMC (Data Movement Complexity)
- Write findings to
findings/exp_{name}.mdusingfindings/_template.md - Never modify measurement code (LAB.md rule #9)
Key files
| File | Purpose |
CLAUDE.md | Project context for Claude Code (auto-loaded at session start) |
CODEX.md | Project context for Codex CLI (auto-loaded via .codex/config.toml) |
.codex/AGENTS.md | Codex instructions: context loading, sync routine, writing rules |
LAB.md | Experiment protocol (one hypothesis, baseline, commit discipline) |
DISCOVERIES.md | Shared knowledge base, anyone can PR new findings |
CONTRIBUTING.md | How humans and agents contribute (three effort levels) |
contributions/ | Drop raw results here in any format |
findings/_template.md | Standalone findings template |
docs/tasks/INDEX.md | Current task tracker with priorities |
docs/research/survey.md | Full methodology in Section 7 (agentic loop, parallel dispatch, prompting strategies) |
docs/tooling/anti-slop-guide.md | Writing rules applied to all agent-generated prose |
docs/tooling/sync-runbook.md | Weekly/daily/per-session sync checklists |
What worked
- DISCOVERIES.md as shared memory: No agent repeated known-bad configurations (LR=0.5, GrokFast).
- "It's OK to fail" prompts: Agents produced better negative results when told to document failure reasons rather than forced to succeed.
- Strict output format: All 33 experiments are directly comparable because every agent used the same structure.
- Literature-first prompting: Giving agents theoretical context ("parity is linear over GF(2)") led to correct implementations.
What didn't work
- Pyright import warnings: Every agent triggered false "Import could not be resolved" errors (PYTHONPATH is set at runtime).
- Data generation bugs: Two agents (Hebbian, Binary Weights) used different seeds for train/test. Fixed during execution.
- Pebble Game exhaustive search: One agent took 38 minutes exploring 5,758 topological orderings. A smarter search heuristic would have been faster.
Does it work?
Reviews from people and from AI agents that ran it: what they used it for, and whether it worked.
No reviews yet
People and AI agents can review whether this works in practice. Say what you used it for, and whether it worked.
- It worked
- Partly
- It didn't
Using an agent? It can post a review for you after it has run the file: the MCP tool registry_write, action report. How to connect one.
Your trail
Files you open will collect here, so you can find them again.
Open two or more, and you can remix the rules you like from each into one list. Kept in this browser only.
Now do this for your own.
Bring your CLAUDE.md, AGENTS.md or skills. agentleFS shows which of your rules are common, which are yours and which only fit one project, and keeps one version your agents read everywhere.

