autoresearch

Autoresearch

Autonomous iteration toward a measurable outcome. The agent modifies code, commits, runs a verify command, keeps improvements, reverts regressions — repeating until a target is hit, a budget is exhausted, or the user interrupts.

Core principles (from Karpathy's autoresearch pattern):

1. One change per iteration
2. Commit before verify
3. Metrics must be mechanical (deterministic, fast, parseable)
4. Keep improvements, revert regressions — no exceptions
5. The branch is sacred — never touch main
6. State survives crashes — resume from last known good
7. Memory spans sessions — what worked/failed carries forward

When to Use

• Task has a measurable numeric metric (size, time, count, score, coverage)
• A verify command exists that outputs the metric deterministically
• "Better" means the number going consistently in one direction
• The agent can make changes autonomously within a defined scope

When NOT to Use

• Subjective goals ("make the UI prettier")
• No verify command available
• Metric requires manual evaluation
• Task needs human judgment per iteration
• Exploratory research without a target

Elicitation Checklist

Before the loop can start, capture these (skip if already provided):

Present as a brief for user approval:

Sage: Autoresearch session configured.

  Goal: [goal statement]
  Metric: [name] ([direction]), target: [target or "none — runs until interrupted"]
  Verify: [command]
  Scope: writable [globs], frozen [globs]
  Budget: [seconds]s per run, [max iterations or "unlimited"]

[A] Start — begin autonomous iteration
[R] Revise — change configuration

The 8-Phase Loop

Each iteration follows 8 phases. Read references/loop-protocol.md for per-phase detail.

Decision rules (Phase 6):

• Exit code ≠ 0 → crash, reset to HEAD
• No METRIC line → crash, reset
• nan/inf → crash, reset
• Metric improved → keep, advance branch
• Metric equal or worse → discard, reset

Runtime Integration

An optional Python runtime handles the deterministic phases (COMMIT, VERIFY, DECIDE, LOG, REPEAT); the agent handles the creative phases (REVIEW, IDEATE, MODIFY). The runtime was extracted from core into its own package in Phase 3 (like sage-memory) — install it with sage add xoai/sage-autoresearch.

Running the runtime — probe first, and degrade LOUDLY if it is absent (announce + log to decisions.md; never silently skip):

if python3 -c 'import autoresearch' 2>/dev/null; then
  python3 -m autoresearch run --brief .sage/work/<slug>/brief.md --project .
else
  echo "Sage: autoresearch runtime not installed — running in degraded (manual)"
  echo "mode. For the deterministic runtime: sage add xoai/sage-autoresearch"
fi

Harness contract: The verify command must print METRIC name=number to stdout. See references/harness-conventions.md.

Session State

All state lives in .sage/work/<YYYYMMDD-slug>/:

Session Resume

On resume (new session, context reset, platform switch):

1. Read autoresearch.md for high-level context
2. Read last 20 lines of autoresearch.jsonl for recent history
3. Verify last JSONL commit matches git log on the branch
4. Continue from next iteration number

See references/session-continuity.md for full protocol.

Memory Integration

Session end: Store a structured summary in sage-memory:

• Winning patterns (what worked)
• Losing patterns (what didn't)
• Best achieved value
• Iteration count

Session start: Search sage-memory for priors on this repo + metric. Inject into IDEATE as "known-good starting points" and "known dead ends."

Quality Gates

Gates are enforced by the runtime, not by prose. The agent cannot bypass them.

References

• references/loop-protocol.md — per-phase inputs, outputs, failure modes
• references/metric-design.md — what makes a good metric
• references/harness-conventions.md — METRIC line contract
• references/stuck-recovery.md — escape local minima
• references/crash-handling.md — retry vs skip decision tree
• references/session-continuity.md — resume protocol