Agent Contracts Strategy
Goal
Transform vague agent instructions into formal "Contracts" that specify deliverables, scope, and validation criteria, mirroring real-world service agreements to ensure reliability.
1. The Contract Data Model
Define tasks using a structured schema to eliminate ambiguity. A valid contract must include:
| Field | Description | Required |
|---|---|---|
| Task Description | Specific, unambiguous description of the objective. | Yes |
| Deliverables & Specs | Precise outcomes and the specific criteria (validators) that make the deliverable acceptable. | Yes |
| Scope | Detailed boundaries of what is included and, critically, what is out of scope. | No |
| Expected Cost/Duration | Constraints on resource usage or time. | Yes |
| Reporting | The feedback loop mechanism (e.g., "Update every 5 steps"). | Yes |
2. The Contract Lifecycle
Agents should not just "execute"; they should negotiate. Implement this lifecycle state machine:
- •Submission: The requester submits a draft contract.
- •Assessment: The agent analyzes feasibility, cost, and ambiguity.
- •Negotiation/Revision:
- •If ambiguous: Agent requests clarification (e.g., "Define 'brief summary'").
- •If too costly: Agent rejects and suggests a narrower scope.
- •Acceptance: Both parties agree on the specs.
- •Execution: The agent generates a plan and executes tasks.
- •Resolution: Deliverables are scored against the initial specifications.
3. Feedback & Iteration
Allow the agent to push back during the Assessment phase using these structured feedback categories:
- •Underspecification: "The requirements for X are missing."
- •Cost Negotiation: "I cannot complete this within the requested token limit."
- •Risk: "This task requires access to restricted data I do not have."
4. Subcontracts (Task Decomposition)
- •Concept: If a contract is too complex, the agent (Contractor) becomes a "Client" and generates Subcontracts for other agents.
- •Mechanism: The agent decomposes the main deliverable into smaller, self-contained contracts, managing the flow and aggregation of results from sub-agents.