More screenshots

    Works with the AI Agents you already use

    Claude CodeClaude CodeCursorCursorCodex CLICodex CLIMuseMuseDotsDots+21 more

    Autoresearch: Autonomous Experiment Loops

    1

    Run a multi-week research project as an agent: literature to hypotheses, locked protocols, fast experiment loops, and synthesis — with the human steering at checkpoints.

    $5

    Secure checkout via Stripe

    Money-back guarantee

    Converts to your local currency at checkout

    Security scanned
    Autoresearch: Autonomous Experiment Loops

    Autoresearch: Autonomous Experiment Loops

    Example session with this skill installed

    Set up a three-week autonomous research project: does retrieval chunk size affect answer accuracy in our support-docs QA bot? We have a 120-question labeled set; the current system uses 800-token chunks. I can review direction once a week.

    • Read your context and instructions
    • Compiled the autoresearch

    PROJECT BOOTSTRAP — chunk size vs. QA accuracy (ready for your checkpoint review)

    State files created: research-state.yaml, research-log.md, findings.md, literature/, experiments/. Budget assumed: 3 weeks, ~15 experiments, weekly direction review.

    Hypotheses (cards written, predictions locked today, before any sweep runs):

    • H1 (mechanistic): chunks under 200 tokens split procedures mid-step, so multi-step answers degrade even if single-fact answers hold.
    • H2 (optimization): 400-token chunks beat the current 800 on exact-match accuracy.
    • H3 (mechanistic): chunk overlap matters more than size for questions whose answers span cross-references.

    Locked evaluation (frozen before experiments):

    • Metric: exact-match accuracy on your 120-question set, split 40 multi-step / 80 single-fact.
    • Baseline to reproduce first: 800-token chunks, expected ~71% overall / ~52% multi-step — if the baseline does not reproduce within tolerance, reproduction becomes experiment zero and nothing else runs.
    • Success: +5 points overall, zero regression on single-fact.

    First inner-loop protocol (written, dated, awaiting your go): sweep 150 / 400 / 800 tokens at fixed overlap. Prediction: multi-step accuracy collapses at 150 (step-splitting), peaks at 400. Outcomes will be labeled confirmatory vs. exploratory in the analysis file either way.

    Outer loop scheduled after every 4–5 experiments and at each weekly review: cluster results, update findings.md, and bring a deepen / broaden / pivot / conclude recommendation with evidence. First progress report: after the size sweep resolves, with the metric trajectory plotted against baseline.

    Nothing runs until you approve the hypotheses and the locked evaluation — that approval is checkpoint one.

    research-state.yaml, research-log.md, findings.md

    Connects securely to your tools. The creator never sees your data.

    About this skill

    Most agent-run research fails the same way: experiments run, results pile up, and nobody can reconstruct what was predicted, what was measured, or why the direction changed. This skill runs the project as a two-loop system with durable state, so the evidence trail survives every session.

    The outer structure: a bootstrap that turns a question into testable hypotheses (survey the literature, find the gaps, distinguish mechanistic hypotheses from optimization ones), then locks the evaluation — metric, baseline, success criteria — before anything runs. The inner loop is fast and disciplined: pick the highest-priority hypothesis, write the protocol with its prediction first, sanity-check the setup, measure the locked metric, and label every outcome confirmatory or exploratory. The outer loop fires every handful of experiments: cluster what worked, ask why, update the findings narrative, and choose deepen, broaden, pivot, or conclude from a decision table — with pivots and conclusions reserved for the human.

    Everything lives in files: a machine-readable state file, a decision log, a findings narrative kept current enough to draft an abstract from, and one directory per experiment holding protocol, raw results, and analysis. Integrity guardrails are built in, not bolted on: protocols before results, baselines must reproduce, negative results recorded as contributions, no metric gaming. A worked example (chunk size vs. QA accuracy) walks a full bootstrap, inner-loop sweep, and outer-loop synthesis. Checklists cover pre-flight per experiment and a quality gate before any progress report ships.

    • It orchestrates research; it does not supply domain expertise or the experiment tooling itself — weak domain tools still produce weak experiments, just well-recorded ones.
    • The discipline has overhead: very small questions (one or two experiments) are faster without it.
    • Checkpoints assume a human who actually reviews; if nobody steers, the loops keep the project honest but cannot choose its direction.
    • Reproducing a baseline is a hard gate — projects whose baselines cannot be reproduced need that fixed before this workflow can measure anything.

    How to use

    1. 1

      Connect your agent to the Agensi MCP

      Copy the prompt and paste it into Claude, ChatGPT, Cursor or any agent that supports MCP. It sets up the connection for you. Manual setup

    2. 2

      Sign in to Agensi

      Your agent shows a sign-in link. Sign in and click Allow.

    3. 3

      Ask for what you need

      Your agent loads this skill and uses it on your task. It always asks before any payment.

    Reviews

    No reviews yet

    Be one of the first to try it. Every listed skill goes through the trust checks below.

    Security scanned

    Checked by our 8-point scan before listing

    Fresh listing

    Recently published to Agensi

    Money-back guarantee

    Doesn't do what its page promises? You get your money back.

    Trust & safety

    Security scan

    Security scan passed 8 days ago

    Permissions

    No special permissions declared or detected

    Listed8 days ago
    Updated6 days ago

    What's inside

    Frequently Asked Questions

    The difference is the paper trail and the locks: hypotheses carry dated predictions, the metric and baseline are frozen before results exist, every outcome is labeled confirmatory or exploratory, and synthesis happens on a schedule instead of at the end. That is what makes the final claims defensible.

    The inner loop can — queued experiments run, get recorded, and feed synthesis without interruption. Direction changes cannot: pivots and conclusions are human checkpoints by design, and the skill will keep executing the current line rather than self-approving a new one.

    That is a result the system is built to honor: each refutation is recorded with what it rules out, the outer loop synthesizes why, and the human decides whether the rigorous negative is the contribution or a pivot is warranted.