SuperAPI AI Directory

polpo-agents

Agents & models

View repository

Updated 10-16 00:02

About this skill

Polpo Agents

Design the smallest agent capability ceiling that supports the intended surfaces. Agent configuration is static policy; request, Channel Route, Loop step, execution, and trusted grants may narrow it but must never widen it.

Workflow

  1. Read the existing agent directory and project settings before changing the definition.
  2. Separate identity and instructions from runtime capabilities. Put instructions in instructions.md; put authored policy in agent.json.
  3. Choose a concrete model or explicit profile. Enable automatic model or execution routing only when the project defines bounded candidates and a deterministic fallback.
  4. Define the global allowedTools ceiling, then add chat or Channel restrictions only where the same agent needs different exposure by surface.
  5. Assign complete skill bundles and Project Loops explicitly. Do not grant every project skill or Loop implicitly.
  6. Configure Memory and sandbox defaults only when the use case needs them. Keep Knowledge sources project-level rather than embedding them into agent instructions.
  7. Validate the materialized agent and every referenced model profile, tool, skill, MCP server, volume, and Loop before deployment.

Invariants

  • Model IDs use provider/model. Profile references use { "profile": "name" }.
  • allowedTools is an authorization ceiling, not a prompt suggestion.
  • toolLoading controls schema disclosure after authorization; it cannot grant a tool.
  • Per-surface and per-execution policies are intersections. Unknown or incompatible tool_choice values fail closed.
  • Chat interactions require both agent permission and client capability.
  • Trusted identity, grants, Connection references, and hidden tool bindings do not belong in model-visible agent configuration.

References