| name | agent-design |
| description | Use when designing a new Flue agent for this repository, including name, runtime placement, instructions, tools, state, deployment target, and tests. |
Agent Design
Use this skill before running bun run agent:new or manually editing Flue agents under agents/*.
Workflow
- Define the agent's job in one sentence.
- Choose runtime placement:
local when it needs local repo/files/tools.cf when it must be externally reachable, scheduled, webhook-driven, shared, or always available.- both only when the same behavior is useful locally and in production.
- Choose state:
- stateless for simple request/response tasks.
- Durable Object-backed state for Cloudflare sessions, schedules, or coordination.
- Draft
instructions as operational behavior, not marketing copy.
- Add one test that proves the agent module exports a Flue agent.
- Run:
bun run agent:new <agent-name> --runtime <node|cloudflare>
bun install
bun run typecheck
bun run check
bun run build
Naming
- Agent directory names are kebab-case under
agents/<name>.
- Package names use
@agent-harness/agent-<name>.
- Runtime mount files use
runtimes/<runtime>/src/agents/<name>.ts.
Avoid
- Do not put deterministic scaffold logic in the agent itself; use scripts/generators.
- Do not deploy to Cloudflare only because it is possible. Use Cloudflare when external availability or durable platform resources matter.
