description: Three-criteria evaluation framework for autonomous agent deployment

Autonomous Agent Deployment

Model Requirement

ALL agents MUST use Opus model. No exceptions, no fallbacks.

Before EVERY Task: Three-Criteria Evaluation

Evaluate these three criteria before starting any task. If any criterion triggers, spawn the appropriate agent(s).

Criterion 1: Research

Question: "Do I need to understand before I change?"

Triggers:

• 3+ files need reading to understand the change
• Working in an unfamiliar area of the codebase
• Change has architectural impact (new patterns, cross-feature, schema changes)
• Existing contracts or docs may be affected but you're not sure which ones

Action: Spawn a research subagent to map the affected area first. The subagent returns a summary of: files involved, contracts affected, patterns in use, risks identified.

Criterion 2: Parallel Work

Question: "Are there independent pieces?"

Triggers:

• Multiple files need changes that don't depend on each other
• Multiple features or components can be built independently
• Tests, docs, and implementation can proceed in parallel

Action: Spawn 2-3 agents for parallel work. Each agent gets a clear scope:

• Define exact files each agent owns (no overlapping writes)
• Define the interface/contract between parallel work items
• Designate one agent as the integrator if work needs merging

Criterion 3: Verification

Question: "Do I need to check my work separately?"

Triggers:

• Changes span 3+ files
• Cross-feature impact (changes in one feature affect another)
• Database schema or migration changes
• Contract or API boundary changes

Action: After implementation, spawn a reviewer subagent that:

• Runs type checking and build verification
• Checks contract/code/doc sync (see contracts-docs rule)
• Validates that no existing patterns were broken
• Reports issues back for fixing before commit

When NOT to Use Agents

Skip the framework entirely when:

• Single-file change with clear, contained scope
• Simple bug fix with an obvious solution
• Task requires fewer than 3 tool calls total
• Context usage is already >60% (spawning agents adds context pressure)
• The user explicitly says "just do it" or similar

Cost Awareness

Agents multiply cost. Use them deliberately, not reflexively.

The cheapest agent is the one you don't spawn. If you can hold the full picture in your current context, just do the work.

Intent Alignment

Before making design decisions, agents MUST check INTENT.md (if it exists) to verify alignment with documented project reasoning. If a task conflicts with documented intent, stop and ask the user before proceeding.

Agent Communication

• Agents share context through files, not through prompt chaining.
• Research agents write findings to a structured summary, not raw dumps.
• Parallel agents must not write to the same file — partition ownership clearly.
• The main thread is responsible for final integration and committing.