Agent Harness¶

The agent harness is OpenCastor's execution engine — it wraps the AI model with safety, cost control, drift detection, and retry logic.

Architecture¶

graph TD
    A[Incoming Command] --> B{P66 Check}
    B -->|preempted| C[ESTOP / Cancel]
    B -->|ok| D[Harness]
    D --> E[Pattern Engine]
    E --> F[LLM Call]
    F --> G{Checks}
    G -->|cost exceeded| H[Abort]
    G -->|drift detected| H
    G -->|ok| I[Execute Action]
    I --> J[Consent Gate]
    J -->|denied| H
    J -->|approved| K[Physical Action]

Harness config schema¶

harness:
  # Core tunables
  max_iterations: 8
  thinking_budget: 2048
  context_budget: 16384
  cost_gate_usd: 0.05
  retry_on_error: true
  drift_detection: true
  p66_consent_threshold: physical

  # Pattern layer (new in v2026.3.21)
  pattern:
    name: single_agent_supervisor   # single_agent_supervisor | initializer_executor | multi_agent
    config: {}

  # Memory layer (new in v2026.3.21)
  memory:
    backend: working                # working | filesystem | firestore
    overflow_strategy: summarize    # summarize | drop_oldest | error
    max_tokens: 8192

  # Security/observability layer (new in v2026.3.21)
  security:
    guardrail: none                 # none | opa
    telemetry: false
    telemetry_endpoint: ""

Pattern engine¶

Three execution patterns are supported:

`single_agent_supervisor`¶

Default. One agent handles the full task loop with a supervisor that checks for drift and budget violations.

`initializer_executor`¶

Two-phase: an initializer decomposes the task and creates a ledger; an executor works through the ledger steps. Good for complex multi-step physical tasks.

`multi_agent`¶

Parallel branches with a coordinator. Useful for server-class hardware with available VRAM.

Memory backends¶

Backend	Use case
`working`	In-memory (default, fastest, no persistence)
`filesystem`	Persists to `/tmp/castor_memory_{session}.json`
`firestore`	Fleet-visible, persists across sessions

Champion configs¶

The research pipeline finds the optimal harness config for your hardware tier. The current fleet champion:

Key	Value
`candidate_id`	`lower_cost`
`cost_gate_usd`	`0.01`
`thinking_budget`	`1024`
`context_budget`	`8192`
`max_iterations`	`6`
`OHB-1 score`	`0.6541` (21/30 tasks)

Apply it: castor harness apply-champion