Agent Quality Standard

Applies to: Every production agent definition
Version: 1.1 | Updated: 2026-07-16

---

Purpose

Agents are programs with probabilistic steps. This standard makes them operable: bounded, observable, and stoppable.

Minimum bar

Definition

• Role is specific with explicit exclusions
• Tool list is least-privilege (no “all tools”)
• Memory requirements documented (what is read/written, retention)
• Communication protocol defined (inputs, outputs, escalation)
• Success termination criteria defined
• Failure termination criteria defined (including max iterations / budget)
• Recovery strategy for each major failure mode

Safety

• Tool execution validated by a harness (schema, authz, policy) — not by model prose alone
• High-impact operations require human approval
• Agent cannot exceed the invoking user’s permissions
• Actions are audited (who/what/when/tool/args redacted appropriately)

Testing

• Happy-path scenarios covered
• At least two adversarial / injection scenarios covered for tool-using agents
• Tool-failure recovery tested
• Loop / budget exhaustion verified

Reference shape

name: example.agent
role: |
  ...
tools: []
memory: []
communication:
  inputs_from: []
  outputs_to: []
  escalates_to: []
termination:
  success: ...
  failure: ...
success_metrics: []

See coder.agent.md for a complete example.

Anti-patterns

Anti-pattern	Fix
Unbounded loops	Hard max iterations + wall-clock budget
“You are helpful” as the whole role	State exclusions and authority limits
Shared static API keys across agents	Per-workload identity
Agent approves its own deploy	Separation of duties

Related

• Agent engineering README
• Multi-agent systems
• Planner agent