Agentic Workflow Canvas — {{title}}
Outcome
- User:
- Goal:
- Deliverable:
- Non-goals:
- Success metrics:
- Unacceptable failures:
Baseline
- Simplest direct approach:
- Current quality / latency / cost:
- Why a workflow may help:
Decomposition
| Step | Input | Output contract | Tool/model | Validator | Failure path |
|---|---|---|---|---|---|
| 1 | | | | | |
State
- Durable workflow state:
- Temporary context:
- Source of truth:
- Resume/idempotency strategy:
Tools and permissions
| Tool | Purpose | Permission | Side effect | Approval required? |
|---|---|---|---|---|
| | | | | |
Evaluation set
- Typical cases:
- Edge cases:
- Known failures:
- Adversarial cases:
- Component metrics:
- Outcome metrics:
Budgets and stopping
- Maximum steps / iterations:
- Time / token / money budget:
- Success stop:
- Failure stop:
- Human escalation:
Risks
- Prompt injection and untrusted content:
- Data privacy:
- Irreversible actions:
- Hallucinated evidence:
- Monitoring and rollback:
Smallest vertical slice
- [ ] One observable end-to-end outcome
- [ ] Eval cases pass
- [ ] Permission boundaries tested
- [ ] Trace is inspectable
- [ ] Cost and latency recorded