Autoresearch

Autoresearch Specialist

Overview

Adapts the interface shape of karpathy/autoresearch into an OSS Agent Lab demonstration. Given a topic, the specialist formats three hypothesis templates, selects a canned finding set, and summarizes its numeric strengths for contract testing.

The current implementation is a contract simulator: it uses fixed hypothesis templates and canned finding sets. It does not search literature, run a model, execute an experiment, or establish facts.

Prerequisites

• Use Python 3.11+ in a local OSS Agent Lab checkout and run pip install -e ..
• Treat every finding and confidence value as synthetic test data.
• Read the runtime contract for accepted methods and limitations.

Capabilities

• research: End-to-end research loop for a given topic or question.
• hypothesis_generation: Produces multiple ranked, testable hypotheses with rationale.
• experiment_design: Selects and runs an experiment method (literature review, simulation, ablation).
• result_analysis: Aggregates findings into insights, a confidence score, and next steps.

Tools

Instructions

1. Extract a non-empty topic and choose literature_review, simulation, or ablation.
2. Run the specialist through the Python API or oss-lab CLI.
3. Label hypotheses, findings, and scores as simulated in any downstream response.
4. Reject requests that require sourced research; use a real research system instead.

Examples

Python API

from agents.specialists.autoresearch.agent import AutoresearchSpecialist
from oss_agent_lab.contracts import Intent, Query, SpecialistRequest

specialist = AutoresearchSpecialist()

request = SpecialistRequest(
    intent=Intent(action="research", domain="science", confidence=0.9),
    query=Query(user_input="effects of sleep deprivation on cognitive performance"),
    specialist_name="autoresearch",
)

result = await specialist.execute(request)
print(result.result["analysis"]["summary"])

CLI

oss-lab run autoresearch "effects of sleep deprivation on cognitive performance"

With method override

request = SpecialistRequest(
    intent=Intent(
        action="research",
        domain="science",
        confidence=0.9,
        parameters={"method": "simulation"},
    ),
    query=Query(user_input="quantum error correction thresholds"),
    specialist_name="autoresearch",
)

Output

{
  "topic": "...",
  "hypotheses": [
    {"id": "h1", "text": "...", "confidence": 0.75, "rationale": "...", "testable": true}
  ],
  "recommended_hypothesis": "h1",
  "experiment": {"id": "abc12345", "method": "literature_review", "status": "completed"},
  "analysis": {
    "summary": "...",
    "key_insights": ["..."],
    "confidence_score": 0.575,
    "next_steps": ["..."]
  }
}

Error Handling

• An unknown method falls back to the canned literature_review finding set; report the effective method and do not imply that sources were consulted.
• Empty findings produce zero confidence and a collect-data recommendation.
• Never convert the returned corpus_scan or meta_analysis labels into citations.

Resources

Wraps karpathy/autoresearch. See the runtime contract for the implemented prototype boundary.