| name | autoresearch |
| description | Autonomous Goal-directed Iteration. Apply Karpathy's autoresearch principles to ANY task. Loops autonomously — modify, verify, keep/discard, repeat. Invoke with /skill:autoresearch or when user says "work autonomously", "iterate until done", "keep improving", or "run overnight". |
| allowed-tools | Bash(git:*) Bash(npm:*) Bash(npx:*) Read Write Edit ask_user show_plan show_research subagent_create_batch dispatch_agent commander_task commander_mailbox show_report |
Autoresearch — Autonomous Goal-directed Iteration
Inspired by Karpathy's autoresearch. Applies constraint-driven autonomous iteration to ANY work — not just ML research.
Core idea: You are an autonomous agent. Modify -> Verify -> Keep/Discard -> Repeat.
When to Activate
- User invokes
/skill:autoresearch or /autoresearch
- User says "work autonomously", "iterate until done", "keep improving", "run overnight"
- Any task requiring repeated iteration cycles with measurable outcomes
Phase 1: Understand (Do This First — Before ANY Work)
Before touching any files, deeply understand the goal. Do NOT rush into iteration.
-
Read relevant files — Scan the codebase to build context around the user's goal. Understand what exists, what patterns are in use, and what's realistic.
-
Identify ambiguities — Based on the goal and codebase context, what's unclear?
- Is the success metric obvious or ambiguous?
- Is the scope (which files to modify) clear?
- Are there constraints the user hasn't mentioned?
- Are there multiple valid interpretations?
-
Ask clarifying questions — If ANY ambiguity exists, use ask_user to ask targeted questions:
ask_user {
question: "I have a few questions before I build the research plan:",
mode: "questions",
options: [
{ label: "1. What metric should define success? (e.g. test coverage %, build time ms, bundle size KB)" },
{ label: "2. Which files/directories are in scope for modification?" },
{ label: "3. Are there any approaches to avoid or constraints I should know about?" },
{ label: "4. What does 'done' look like — a specific target, or iterate until interrupted?" }
]
}
Tailor questions to the specific goal. Don't ask about what's already clear. Ask about genuine ambiguities.
-
Skip if crystal clear — If the goal is unambiguous (clear metric, scope, exit criteria), skip questions and proceed to Phase 2. State briefly why no questions are needed.
-
Synthesize understanding — Form a concrete statement: Goal, Metric (what + direction + verify command), Scope (in/out), Constraints, Exit criteria.
-
Save research session — Create .context/research-sessions/<session-id>.json with the initial session data: goal, metric, scope, clarifying Q&A, status "understanding". Store the session_id for updates throughout the lifecycle.
Phase 2: Plan (Present Before Executing)
Now write and present a research plan for user approval. Do NOT start iterating without approval.
-
Establish baseline — Run the verification command to get a starting metric value.
-
Write the research plan — Create .context/autoresearch-plan.md:
# Autoresearch Plan: <goal summary>
## Goal
<Concrete goal statement>
## Metric
- **Measuring:** <what>
- **Direction:** <higher/lower is better>
- **Verify command:** `<command>`
- **Baseline:** <current value>
- **Target:** <target value or "continuous improvement">
## Scope
- **In scope:** <files/directories to modify>
- **Read only:** <files for context only>
- **Out of scope:** <excluded areas>
## Strategy
Ordered approaches, most to least promising:
1. <First approach — why promising>
2. <Second approach — what it explores>
3. <Third approach — alternative angle>
4. <Fourth approach — radical idea>
5. <Fifth approach — simplification play>
## Iteration Plan
- **Mode:** <bounded (N) / unbounded>
- **Estimated time per iteration:** <seconds/minutes>
- **When stuck:** Re-read plan, combine near-misses, try opposites
## Exit Criteria
- <When to stop>
-
Present for approval:
show_plan { file_path: ".context/autoresearch-plan.md", title: "Autoresearch Plan: <goal>" }
- Approved → proceed to Phase 3
- Declined → revise based on feedback and re-present
-
Update session — Set status to "planning", save plan content and baseline metric.
Phase 3: Setup & Begin
With understanding confirmed and plan approved, set up tracking and start.
- Create results log — Create
autoresearch-results.tsv (see references/results-logging.md)
- Record baseline — Log the baseline metric from Phase 2 as iteration #0
- Commander tracking — If available, create task group and broadcast start (see Commander Integration below)
- Update session — Set status to "researching"
- Begin the loop — Start iterating immediately. No further confirmation needed.
The Loop
Read references/autonomous-loop-protocol.md for full protocol details.
LOOP (FOREVER or N times):
1. Review: Read current state + git history + results log
2. Ideate: Pick next change based on goal, past results, what hasn't been tried
3. Modify: Make ONE focused change to in-scope files
4. Commit: Git commit the change (before verification)
5. Verify: Run the mechanical metric (tests, build, benchmark, etc.)
6. Decide:
- IMPROVED -> Keep commit, log "keep", advance
- SAME/WORSE -> Git revert, log "discard"
- CRASHED -> Try to fix (max 3 attempts), else log "crash" and move on
7. Log: Record result in results log
8. Repeat: Go to step 1.
- If unbounded: NEVER STOP. NEVER ASK "should I continue?"
- If bounded (N): Stop after N iterations, print final summary
Critical Rules
- Loop until done — Unbounded: loop until interrupted. Bounded: loop N times then summarize.
- Read before write — Always understand full context before modifying
- One change per iteration — Atomic changes. If it breaks, you know exactly why
- Mechanical verification only — No subjective "looks good". Use metrics
- Automatic rollback — Failed changes revert instantly. No debates
- Simplicity wins — Equal results + less code = KEEP. Tiny improvement + ugly complexity = DISCARD
- Git is memory — Every kept change committed. Agent reads history to learn patterns
- When stuck, think harder — Re-read files, re-read goal AND
.context/autoresearch-plan.md for planned strategy, try next untried approach from the plan, combine near-misses, try radical changes. Don't ask for help unless truly blocked by missing access/permissions
Principles Reference
See references/core-principles.md for the 7 generalizable principles from autoresearch.
Commander Integration (Task Tracking & Visibility)
When Commander is available, autoresearch MUST track every iteration as a Commander task. This gives the dashboard full visibility into autonomous work — just like the tasks extension does for manual workflows.
Setup Phase (Phase 3) — Create Task Group
After establishing the baseline and getting plan approval, create a Commander task group for this research session:
commander_task {
operation: "group:create",
group_name: "Autoresearch: <goal summary>",
initiative_summary: "<full goal description with metric and scope>",
total_waves: 1,
working_directory: "<cwd>",
tasks: []
}
Store the returned group_id — all iteration tasks will be added to this group.
Send an initial mailbox status broadcast:
commander_mailbox {
operation: "send",
from_agent: "autoresearch",
to_agent: "commander",
body: "Autoresearch started: <goal>. Baseline metric: <value>. Scope: <files>. Plan approved.",
message_type: "status"
}
Per-Iteration — Create → Claim → Complete
Before modifying (step 3 of each loop iteration), create and claim a Commander task:
commander_task { operation: "create", description: "Iteration #N: <planned change>", working_directory: "<cwd>", group_id: <group_id> }
commander_task { operation: "claim", task_id: <task_id>, agent_name: "autoresearch" }
After logging results (step 7 of each loop iteration), complete the task with the outcome:
commander_task { operation: "complete", task_id: <task_id>, result: "<status>: <description>. Metric: <old> → <new> (delta: <delta>)" }
Also add a comment to the task with detailed results:
commander_task { operation: "comment:add", task_id: <task_id>, body: "Status: <keep|discard|crash>\nMetric: <value> (delta: <delta>)\nCommit: <hash or '-'>\nDescription: <what was tried>", agent_name: "autoresearch" }
Note: Use complete for ALL outcomes (keep, discard, crash). Discards and crashes are expected in autoresearch — they're not failures. Reserve fail only for unrecoverable errors that halt the entire loop.
Status Broadcasts — Every ~5 Iterations
Every 5 iterations, send a mailbox status update AND add a comment to the group:
commander_mailbox {
operation: "send",
from_agent: "autoresearch",
to_agent: "commander",
body: "Autoresearch progress — Iteration #N: metric at <value> (baseline: <baseline>). Keeps: X | Discards: Y | Crashes: Z",
message_type: "status"
}
Research Complete — Report & Implementation Handoff (MANDATORY)
When the loop ends (bounded mode reaching N, or goal achieved):
- Final mailbox broadcast with full summary:
commander_mailbox {
operation: "send",
from_agent: "autoresearch",
to_agent: "commander",
body: "Autoresearch complete (N iterations). Baseline: <X> → Final: <Y> (delta: <Z>). Keeps: A | Discards: B | Crashes: C. Best iteration: #M — <description>",
message_type: "result"
}
-
Compile findings & next steps — Extract prioritized, actionable implementation items from the research. Update the session file with findings, next steps array, and final metric.
-
Research report — Present via show_report framed as a handoff:
show_report {
title: "Research Complete — Ready for Implementation: <goal>",
summary: "## Research Results\n\n...\n\n## Prioritized Next Steps\n\n1. <action item>\n2. ...\n\n## Recommended Implementation Approach\n\n<how to implement>"
}
-
Ask about implementation — Use ask_user to offer three choices:
- Implement now → spawn a team of builder agents via
subagent_create_batch
- Save & pause → set session to "paused", resume later via
/research
- Done → mark session "complete"
-
Implementation (if chosen) — Update session to "implementing", create Commander task group, dispatch builders, track completion. When done, present final comprehensive report covering research results AND implementation work. Set session to "complete".
-
Preserve the plan — Leave .context/autoresearch-plan.md intact. Leave the session file for browsing via /research.
Graceful Degradation
All Commander calls are optional. If Commander is unavailable:
- Skip
commander_task and commander_mailbox calls silently
- The local
autoresearch-results.tsv log remains the primary record
- The
show_report call still works (it only needs git, not Commander)
- Never let a Commander error interrupt the autonomous loop
Adapting to Different Domains
| Domain | Metric | Scope | Verify Command |
|---|
| Backend code | Tests pass + coverage % | src/**/*.ts | npm test |
| Frontend UI | Lighthouse score | src/components/** | npx lighthouse |
| ML training | val_bpb / loss | train.py | uv run train.py |
| Blog/content | Word count + readability | content/*.md | Custom script |
| Performance | Benchmark time (ms) | Target files | npm run bench |
| Refactoring | Tests pass + LOC reduced | Target module | npm test && wc -l |
Adapt the loop to your domain. The PRINCIPLES are universal; the METRICS are domain-specific.
Session Persistence
Every autoresearch session is saved to .context/research-sessions/<session-id>.json. This enables:
- Resume later — pick up where you left off via
/research command
- Browse history — see all past research sessions in the research browser
- Track lifecycle — from understanding through implementation completion
Update the session file at every major transition: understand → plan → research → implement → complete. On every "keep" iteration or every ~5 iterations, append iteration data to the session. This creates a complete record of the research lifecycle that can be browsed and resumed.