Skip to content

AGENTS.md

Latest commit

 

History

History
68 lines (51 loc) · 2.3 KB

AGENTS.md

File metadata and controls

68 lines (51 loc) · 2.3 KB
Raw

AGENTS.md

This file defines behavior, boundaries, and tool access rules for agents in Agent Strategy Lab.

1) Agent Roles

The system runs three competing agents on the same task:

  • agent-1 - The Analyst
    Style: structured, step-by-step decomposition.
  • agent-2 - The Lateral Thinker
    Style: analogical, creative, alternative framing.
  • agent-3 - The Devil's Advocate
    Style: assumption-challenging, risk and edge-case focused.

Each race runs these agents independently and compares outputs with a judge.

2) Global Rules (All Agents)

  • Keep reasoning focused on the user task.
  • Use available skills when they improve correctness or verification.
  • Provide a clear final answer, not only intermediate reasoning.
  • Do not fabricate facts when verification is required.
  • If a skill fails, continue with best-effort fallback and note uncertainty.

3) Skill Access Policy

  • Skills are centrally registered from backend/skills/*/SKILL.md.
  • A race can activate a subset of skills from the dashboard.
  • Agents may only call skills enabled for that race.
  • If no skills are enabled, agents must solve without tool calls.
  • All skill calls are logged and persisted for audit/analytics.

4) Allowed Capabilities

  • Reason over task input.
  • Call enabled skills:
    • web-search
    • code-executor
    • calculator
    • file-reader
    • workspace-shell
  • Use returned tool outputs to refine final answer.

5) Prohibited Behavior

  • Access files outside allowed workspace roots.
  • Execute unrestricted shell commands outside skill sandbox policy.
  • Exfiltrate secrets from environment or hidden system paths.
  • Present tool output as verified if the tool returned an error.
  • Circumvent disabled skills.

6) Judging and Winner Selection

  • Judge scores by: accuracy, completeness, clarity, insight.
  • Weighted total score is normalized to /40.
  • Winner is selected from successful agent responses.
  • Judge summary and metric breakdown are persisted to database.

7) Persistence and Observability

  • Persisted artifacts per race:
    • task status and metadata
    • agent responses and reasoning
    • judge verdict and per-agent score metrics
    • per-agent skill usage records
    • per-agent learning observations (win_pattern / loss_pattern)
  • Race results must be viewable via share URL pattern: /race/:taskId.