Skip to content
AletheIA Docs
Esc
↑↓navigate↵open⌘Jpreview
On this page

Agent Harness Contract

What this is

An Agent Harness Contract (AHC) is the declared operating envelope for a task, written before the task runs. A skill says how to think; a knowledge pack says what content is authorized; an AHC says in what environment the agent may act — with which tools, autonomy, gates, sensors, logs, and rollback.

It is a contract, not a runtime, orchestrator, or policy engine, and deliberately not a new “layer”. It is docs-first and provider-agnostic.

Where it sits

The AHC composes three governance surfaces the framework already has. Each governs a distinct question at its own granularity:

User / Request
  → AletheIA task contract
    → Knowledge Governance (authorized context)
      → Agent Harness Contract  ← the declared envelope (this document)
        → Adaptive skill selection
          → Runtime (Codex / Claude Code / …)
            → Tools + Sensors + Logs
              → Evidence + Decision Record + Handoff
REGC (how much effort, per-slice)
  → AHC (the declared envelope, per-task)
    → AHGE (execution + record, per-action)
      → Knowledge Governance (authorized context)

AHC declares the envelope; AHGE records each action within it; REGC sets the effort; Knowledge Governance decides what context is authorized. The normative field-by-field spec and the mapping to these surfaces live in the Agent Harness Contract spec.

Boundary table

Surface Decides AHC’s relationship
Skill how to execute a capability AHC lists allowed_skills / blocked_skills; it does not define skill behavior
Knowledge Governance source authority, sensitivity, precedence AHC references allowed_knowledge_packs and retrieval_mode; it never decides authority
REGC effort level, escalation, stop, quality floor AHC inherits the effort; it does not re-decide it
AHGE per-action permission, draft/commit, budgets, trace AHC declares the envelope AHGE enforces and records per action

Principles

  1. Harness is not skill. Skill orients behavior; harness defines environment, tools, sensors, gates, and limits.
  2. Harness is not Knowledge Governance. Knowledge Governance controls authorized context; the harness controls execution conditions.
  3. Control proportional to risk. Reversible, local tasks may use a light harness; structural, sensitive, or irreversible tasks require more gates.
  4. Sensor before judgment. Before asking another model for a semantic opinion, use computational sensors when available: tests, linters, type checkers, scripts, structural validation.
  5. Context needs hygiene. Long sessions require offloading, checkpoints, recorded decisions, and re-anchoring — see context-rot-controls.md.
  6. A harness expires. Controls, guides, prompts, scripts, and sensors must be reviewed. A control that never fires may be excellent or useless; it needs evaluation — see harness-expiration-review-checklist.md.

Was this page helpful?