A practical cheat sheet for designing robust, testable, future-proof agent workflows.
This is optimized for:
- single main agent ownership
- explicit skill usage
- sequential, deterministic execution
- “good enough” optimization
- One main agent owns intent and decisions
- No role-based discussions (“designer vs architect”)
- No swarm-style negotiation loops
Coordination happens in the workflow, not in conversation.
Reason: Role-based multi-agent choreography is token-heavy, less deterministic and often doesnt't survive model upgrades.
Think in phases with clear responsibilities.
Each phase defines:
- goal
- allowed skills/tools
- expected artifact
- exit condition
- Skills are named and invoked
- Never “guess” a capability
- Prefer “use X skill” over “act as Y role”
Each phase produces something concrete:
- file
- diff
- screenshot
- test output
- image
If there is no artifact, the phase is probably unclear.
# Agent Workflow
## Constraints
- Follow phases sequentially
- Do not skip phases
- Do not invent tools or skills
- Produce artifacts for each phase
---
## Phase 1: Plan
**Goal:** Define the approach
**Allowed skills:** reasoning only
**Output:** `plan.md`
**Exit condition:** clear steps, no implementation
---
## Phase 2: Implement
**Goal:** Write code according to plan
**Allowed skills:** filesystem, shell
**Output:** code + tests
**Exit condition:** build succeeds
---
## Phase 3: Validate
**Goal:** Verify behavior
**Allowed skills:** test runner / browser automation
**Output:** test results, logs, screenshots
**Exit condition:** tests pass or failures explained
---
## Phase 4: Design Asset (conditional)
**Goal:** Produce visual artifact
**Allowed skills:** image generation
**Output:** image file (e.g. `logo.svg`)
**Exit condition:** image matches constraints
---
## Phase 5: Finalize
**Goal:** Prepare final state
**Allowed skills:** git (commit only)
**Output:** commit
**Exit condition:** clean working tree