Chapter 6: The Agent Protocol

What Problem Does This Solve?

An LLM given a vague instruction like "improve this ML training script" will:

1. Make a change
2. Ask "should I test this?"
3. Wait for human response
4. Make another change
5. Ask "does this look good?"
6. Never stop asking

autoresearch solves this by encoding a complete, unambiguous protocol in program.md. The document specifies:

• Exactly how to branch the repository
• Exactly how to run the experiment
• Exactly how to measure success
• Exactly what to do when it fails
• Exactly how to log results
• That the agent should NEVER ask the human anything

The result is an LLM that behaves like a specialized, self-directing research engineer — not because it was fine-tuned for this task, but because its instructions are complete enough to leave no gaps requiring human input.

program.md as "Research Org Code"

The name "research org code" is deliberate. In a human research organization:

• A lab has a protocol (how experiments are run)
• A lab has standards (what constitutes a valid result)
• A lab has a culture (what kinds of discoveries are valued)
• A lab has autonomy norms (when to escalate vs proceed independently)

program.md encodes all four for a single-agent "research organization." It is the agent's entire institutional context.

graph TD
    PM[program.md] --> PROTOCOL[Experimental Protocol<br/>branch → modify → commit → run → measure]
    PM --> STANDARDS[Quality Standards<br/>val_bpb must improve, memory must not explode]
    PM --> CULTURE[Research Culture<br/>simplicity criterion, prefer deletion over addition]
    PM --> AUTONOMY[Autonomy Norms<br/>NEVER STOP, NEVER ASK, human is asleep]
    PM --> LEDGER[Record-Keeping<br/>results.tsv format, git history as ground truth]

The Branch Naming Convention

The first thing the agent does when starting a session is create a branch:

git checkout -b autoresearch/<descriptive-tag>

The <descriptive-tag> should describe the agent's planned exploration direction:

autoresearch/rope-scaling-experiments
autoresearch/deeper-narrower-architecture
autoresearch/muon-warmup-variants
autoresearch/sliding-window-ablations

This naming convention serves multiple purposes:

1. Isolation: experiments on different branches do not interfere with each other
2. Discoverability: a human reviewing the repository can see what directions were explored
3. Parallelism: multiple agents can run simultaneously on different branches without conflicts
4. Cleanup: git branch -D autoresearch/* removes all agent branches cleanly

The Experiment Loop

The core protocol is a tight loop:

LOOP FOREVER:
  1. Hypothesize: choose one modification to train.py
  2. Implement: edit train.py
  3. Commit: git commit -am "<description>"
  4. Run: uv run train.py > run.log 2>&1
  5. Measure: grep "val_bpb=" run.log | tail -1
  6. Decide:
     - If val_bpb improved (lower): keep commit, append to results.tsv
     - If val_bpb did not improve: git reset --hard HEAD~1
  7. Go to step 1

flowchart TD
    START[Start session] --> BRANCH[git checkout -b autoresearch/tag]
    BRANCH --> HYPOTHESIZE[Choose one modification to train.py]
    HYPOTHESIZE --> IMPLEMENT[Edit train.py]
    IMPLEMENT --> COMMIT[git commit -am description]
    COMMIT --> RUN[uv run train.py > run.log 2>&1]
    RUN --> GREP[grep val_bpb= run.log]
    GREP --> COMPARE{val_bpb < current best?}

    COMPARE -->|Yes: improved| LOG[Append to results.tsv]
    LOG --> UPDATE[Update current best]
    UPDATE --> HYPOTHESIZE

    COMPARE -->|No: not improved| RESET[git reset --hard HEAD~1]
    RESET --> HYPOTHESIZE

    RUN -->|exit code 1 FAST_FAIL| FAST_FAIL[Log as failed, git reset]
    FAST_FAIL --> HYPOTHESIZE

Git as the Experiment Ledger

The decision to use git as the experiment tracking system is elegant in its simplicity:

Keeping an Improvement

When val_bpb improves, the commit stays:

# The commit already exists from step 3
# Nothing to do — the state is preserved in git history

Rejecting a Failure

When val_bpb does not improve:

git reset --hard HEAD~1

This rolls back to the previous state: the modification to train.py is undone, the commit is removed from history. The repository is exactly as it was before the failed experiment.

gitGraph
    commit id: "baseline: val_bpb=1.8342"
    commit id: "try: RoPE scaling → 1.8201" type: HIGHLIGHT
    commit id: "try: wider MLP (rejected)" type: REVERSE
    commit id: "try: fewer KV heads → 1.8150" type: HIGHLIGHT
    commit id: "try: squared ReLU off (rejected)" type: REVERSE
    commit id: "try: longer warmup → 1.8089" type: HIGHLIGHT

Note: after git reset --hard HEAD~1, the "rejected" commits disappear from the history. What the agent actually sees in git log is a clean sequence of improvements. The rejections only appear in results.tsv (as rows with status=rejected).

Why Not a Database or MLflow?

The git approach has several advantages over dedicated experiment tracking systems:

Property	git reset	MLflow / W&B
Zero infrastructure	Yes	Requires server or account
Automatic versioning	Yes	Manual
Rollback built-in	Yes	Requires custom logic
Reproducibility	Exact (commit hash)	Depends on artifact storage
Offline capable	Yes	Usually not
Human-readable	Yes	Requires UI

The tradeoff is that git does not store all the failed experiments' code — only results.tsv records that they were attempted. If you want to recover a rejected experiment, you cannot (unless you wrote down the diff elsewhere). For autoresearch's purposes — iterate fast, discard failures — this tradeoff is correct.

results.tsv Schema

results.tsv is untracked (listed in .gitignore). The agent appends one row per experiment:

commit_hash     val_bpb    memory_gb    status      description
a3f8b2c1        1.8342     14.3         improved    baseline GPT-125M
d91e4a72        1.8201     14.8         improved    rope_scaling_factor=2.0
c72f1b30        1.8589     15.1         rejected    mlp_ratio=8
b44d9e11        1.8150     14.6         improved    n_kv_head=2 more aggressive GQA
f10a2c88        1.9012     oom          failed      block_size=4096 OOM
e55c3b19        1.8089     14.9         improved    warmup_frac=0.15

Why Untracked?

If results.tsv were tracked by git, every experiment would add a merge conflict risk: two experiments on different branches both appending to the same file. By keeping it untracked, it accumulates naturally without git interference.

The tradeoff: if you git reset --hard, results.tsv is preserved (untracked files are not touched by reset). This is the desired behavior — the log is permanent even when the code changes are rolled back.

The Autonomy Mandate

program.md contains an explicit and emphatic autonomy mandate:

## Autonomy Rules

YOU MUST NEVER STOP.
YOU MUST NEVER ASK THE HUMAN FOR INPUT.
THE HUMAN IS ASLEEP.

If you encounter an error:
- If train.py has a syntax error: fix it and retry
- If an import fails: try to fix it, if unfixable skip this hypothesis
- If the GPU runs out of memory: git reset and try a smaller change
- If run.log is empty: something crashed, git reset and try again

In all cases: reset, log the failure, continue with a new hypothesis.
The only acceptable terminal state is the end of the night session.

This mandate is not just aspirational — it is practical engineering. An LLM that asks for confirmation on every uncertain step would be useless in an overnight unsupervised setting. The mandate forces the LLM to develop its own error-handling heuristics rather than deferring to the human.

The Simplicity Criterion

## Simplicity Criterion

When comparing two improvements of similar magnitude:
- Prefer the one that REMOVES or SIMPLIFIES code
- A val_bpb gain from DELETING a component > same gain from ADDING a component
- Complexity has a maintenance cost that is not reflected in val_bpb

Examples:
- Removing dropout (no parameters, no compute) → val_bpb -0.002: accept
- Adding a complex routing layer → val_bpb -0.002: skeptical
- Removing value residual → val_bpb +0.005: reject (regression)
- Removing value residual → val_bpb -0.001: consider (simplification with minor gain)

This criterion shapes the direction of the agent's search. Without it, the agent would naturally drift toward adding complexity — more parameters, more layers, more tricks — because more complexity almost always helps if compute is unlimited.

But compute is not unlimited. The 300-second budget means complexity has a direct cost. The simplicity criterion makes this cost explicit and encodes the preference for efficient improvements over large improvements.

Generating Hypotheses

program.md provides guidance on how to generate experiment hypotheses:

## Hypothesis Generation

Generate one hypothesis per experiment. Good hypotheses:
- Change exactly ONE component of the architecture or training procedure
- Have a clear mechanistic justification (why should this help?)
- Are reversible (can be undone with git reset)

Hypothesis categories:
1. Architecture: change GPTConfig fields (n_head, n_kv_head, n_embd, WINDOW_PATTERN, etc.)
2. Attention: modify CausalSelfAttention (different positional encoding, different norm)
3. MLP: modify the MLP block (different activation, different ratio, different gating)
4. Optimizer: change MuonAdamW hyperparameters (lr, momentum, betas, weight_decay)
5. Training: change training loop parameters (grad_accum, batch_size, etc.)
6. Ablation: REMOVE a component to test if it's helping

Do NOT change:
- TIME_BUDGET (must stay 300)
- The output format (val_bpb=X.XXXX | memory_gb=XX.X | steps=NNNN)
- Any import from prepare.py
- prepare.py itself

Interacting with the Agent

The agent is invoked by passing program.md as a system prompt to an LLM that has tool-use capability (shell commands, file editing):

Using Claude

# Open Claude Code in the autoresearch directory
cd autoresearch
claude  # starts Claude Code session

# Then paste or type:
# "Read program.md and begin the autoresearch protocol.
#  Current baseline is val_bpb=1.8342 from commit a3f8b2c1.
#  Go."

Using the API (Headless)

import anthropic

client = anthropic.Anthropic()
program_md = open("program.md").read()
baseline_context = "Current best val_bpb=1.8342 (commit a3f8b2c1). Begin experiments."

response = client.messages.create(
    model="claude-opus-4-5",
    max_tokens=8192,
    system=program_md,
    messages=[{"role": "user", "content": baseline_context}],
    tools=[...],  # shell_exec, file_write, file_read tools
)

The agent uses shell execution tools to run git, uv run, and grep commands, and file editing tools to modify train.py.

Error Handling Protocol

The protocol specifies how to handle each class of error:

graph TD
    ERROR[Error during experiment] --> TYPE{Error type?}

    TYPE -->|Syntax error in train.py| FIX_SYNTAX[Fix the syntax error\nretry same hypothesis]
    TYPE -->|Import error| FIX_IMPORT[Try to fix\nif unfixable, skip and try new hypothesis]
    TYPE -->|OOM GPU error| OOM[git reset\nlog: status=failed, description=OOM\ntry smaller modification]
    TYPE -->|NaN loss fast-fail| NAN[git reset\nlog: status=failed, description=NaN\nreturn to safer region]
    TYPE -->|Empty run.log| CRASH[git reset\nlog: status=failed, description=crash\nnote the hypothesis for later analysis]
    TYPE -->|No val_bpb in run.log| PARTIAL[git reset\nlog: status=failed, description=incomplete run]

    FIX_SYNTAX --> RETRY[Retry]
    FIX_IMPORT --> NEXT[Next hypothesis]
    OOM --> NEXT
    NAN --> NEXT
    CRASH --> NEXT
    PARTIAL --> NEXT

Session Boundaries and Resumption

program.md specifies how to resume after a session ends (e.g., GPU time expired, network disconnection):

## Session Resumption

When resuming an existing session:
1. git log --oneline -20 to see recent history
2. cat results.tsv | tail -20 to see recent experiments
3. Find the current best val_bpb from results.tsv
4. Resume from the current HEAD (do not re-run old experiments)
5. Continue the experiment loop from step 1

Do NOT create a new branch — continue on the existing autoresearch/<tag> branch.

Chapter Summary

Component	Purpose	Key Detail
Branch naming	autoresearch/<tag>	Isolates experiment directions, enables multi-agent
Experiment loop	modify → commit → run → measure → keep/reset	~8 minutes per cycle
git as ledger	commits for improvements, reset for failures	Zero extra infrastructure
results.tsv	Untracked experiment log	Preserved through git reset
Autonomy mandate	NEVER STOP, NEVER ASK	Handles all errors independently
Simplicity criterion	Prefer deletion over addition	Shapes search toward efficient improvements
Hypothesis generation	One change per experiment	Controls for confounds
Error handling	Class-specific recovery procedures	No dead-ends for the agent

In the next chapter, we examine analysis.ipynb — the Jupyter notebook for reading results.tsv, visualizing the overnight progress curve, identifying the best experiments, and extracting patterns from 100 experiment runs.