Agent contract (spec)

Agents are workers. They do not orchestrate execution.

Responsibilities

Agents MAY:

• read/transform text
• produce structured artifacts
• emit confidence scores

Agents MUST NOT:

• manage pipeline lifecycle or transitions
• persist state across runs (unless explicitly designed and documented)
• write final artifacts directly (that is FINALIZE responsibility)

Input contract

Every agent run should be representable as an AgentInputSchema:

• task_goal (string)
• context_id (string)
• payload (mapping)
• metadata (mapping)

Output contract

Agent outputs must satisfy AgentOutputSchema:

• text (non-empty)
• confidence in [0, 1]
• metadata.contract_version matching the runtime contract version

This strictness is deliberate: it keeps downstream decisions stable even as prompts or models evolve.

(Implementation reference: src/bijux_agent/models/contract.py.)