| name | ai-agent-implementation-workflow |
| description | Implement agentic behaviors safely with clear tool boundaries, deterministic contracts, and incremental verification across prompts, tools, and orchestration code. Use when this capability is needed. |
| metadata | {"author":"Bryan-Roe"} |
AI Agent Implementation Workflow
What This Skill Produces
Use this skill to implement or update agent behavior with stable contracts. The expected result is:
- explicit agent goal and decision boundaries
- predictable tool invocation flow
- schema-safe outputs for downstream consumers
- graceful fallback when tools/providers fail
- targeted tests or smoke checks proving behavior
When to Use
Use this skill when you need to:
- add a new agent workflow
- refine tool usage logic for an existing agent
- fix agent output/schema instability
- harden retry/fallback behavior in agent loops
- align agent behavior across CLI/API/UI surfaces
Common trigger phrases:
- "implement this agent behavior"
- "add tool-calling to the agent"
- "fix unstable agent responses"
- "make the agent robust"
- "agent output schema keeps breaking"
Procedure
-
Define contract first
- Lock input/output schema and required fields.
- Clarify what is best-effort vs required behavior.
-
Constrain tool boundaries
- List which tools can be called and for what reasons.
- Keep side-effecting actions explicit and auditable.
-
Implement minimal orchestration
- Prefer small deterministic control flow over deep branching.
- Make retries bounded and reason-aware.
-
Handle degraded mode intentionally
- Return actionable errors when hard requirements are missing.
- Use safe fallback only when it preserves contract meaning.
-
Verify behavior incrementally
- Add focused tests/smokes for primary path + fallback path.
- Confirm output schema is stable across paths.
-
Validate integration surface
- Ensure consuming endpoints/UI can parse new outputs.
- Avoid silent breaking changes in event/JSON structure.
Quality Checks
Before finishing, confirm that:
- output schema is deterministic and documented
- tool usage boundaries are explicit
- retries/fallbacks are bounded and observable
- failures are actionable, not silent
- integration consumers remain compatible
Source: Bryan-Roe/Aria — distributed by TomeVault.