AI Agents artefact templates and rubrics

How to use this

Practical, not perfection

Pick the tier you are studying and copy the template into your notes.
Complete it for one real example (work, a side project, or a safe fictional scenario).
Use the rubric to tighten it until it is defensible.
Paste a short reflection into your CPD record: what you assumed, what evidence you would keep, and what you would do next.

Foundations. Agent capability scope memo

Template + rubric + example

Template

Title: Agent capability scope memo

1) What this agent is designed to do (one sentence)
- 

2) What it must NOT do (three boundaries)
- 
- 
- 

3) Tools and permissions
- Available tools: 
- Permission boundaries: 

4) Context and memory constraints
- Context window limit: 
- What is persisted: 
- What is ephemeral: 

5) Fallback and escalation
- When does the agent stop and escalate?
- How do users intervene?

6) Evidence to keep
-

Rubric (what “good enough” looks like)

Scope and boundaries are clear (what is in/out).
Tool permissions are explicit and minimal.
Escalation triggers are clear and testable.
Evidence is identified (what you would log/measure/keep).

Worked example (short)

Example (short):
This agent schedules meetings by reading calendars and sending invites.
It must not access email content, delete events, or share availability externally.
Fallback: if no overlapping time is found, summarise conflicts and ask for guidance.

Core concepts. Agent design patterns checklist

Template + rubric + example

Template

Title: Agent design patterns checklist

1) Pattern selected
- [ ] ReAct (reason + act loop)
- [ ] Chain of thought
- [ ] Tool use orchestration
- [ ] Multi-agent delegation
- [ ] Other: 

2) Memory architecture
- Short-term (in-context): 
- Long-term (retrieval): 
- Structured (database): 

3) Tool calling strategy
- Tool selection method: 
- Error handling: 
- Retry policy: 

4) Observation and feedback loop
- What signals indicate success?
- What triggers re-planning?

5) Evidence to keep
-

Rubric (what “good enough” looks like)

Assumptions are explicit and testable.
Pattern selection is justified by the task requirements.
Memory boundaries prevent context pollution.
Evidence is identified (what you would log/measure/keep).

Worked example (short)

Example (short):
Pattern: ReAct with tool-use for structured actions.
Memory: rolling window of last 5 exchanges + vector retrieval for policies.

Practical building. Agent implementation specification

Template + rubric + example

Template

Title: Agent implementation specification

1) Runtime environment
- Host: 
- Model: 
- Framework: 

2) Tool definitions
| Tool name | Input schema | Output schema | Side effects |
|-----------|--------------|---------------|--------------|
|           |              |               |              |

3) Integration points
- APIs consumed: 
- Events emitted: 
- Human-in-the-loop checkpoints: 

4) Testing strategy
- Unit tests for tools: 
- Integration tests for workflows: 
- Adversarial inputs to test: 

5) Evidence to keep
-

Rubric (what “good enough” looks like)

Tool schemas are precise and validated.
Side effects are identified and reversible where possible.
Evidence is identified (what you would log/measure/keep).
Next actions are concrete and realistic.

Worked example (short)

Example (short):
Framework: LangGraph with OpenAI function calling.
Human checkpoint: confirm before sending external communications.

Security and ethics. Agent threat model and guardrails

Template + rubric + example

Template

Title: Agent threat model and guardrails

1) Threat scenarios
| Threat | Impact | Likelihood | Mitigation |
|--------|--------|------------|------------|
| Prompt injection | | | |
| Tool misuse | | | |
| Data exfiltration | | | |
| Resource exhaustion | | | |

2) Input validation
- User input sanitisation: 
- Context injection protections: 

3) Output guardrails
- Content filtering: 
- Action rate limits: 
- Irreversible action approvals: 

4) Ethical boundaries
- What decisions should never be automated?
- How is bias monitored?

5) Incident response
- Detection signals: 
- Response playbook: 

6) Evidence to keep
-

Rubric (what “good enough” looks like)

Prompt injection mitigations are explicit.
High-impact actions require human approval.
Evidence is identified (what you would log/measure/keep).
Next actions are concrete and realistic.

Worked example (short)

Example (short):
Threat: prompt injection via user-supplied context.
Mitigation: separate system and user instructions, use structured outputs, log all tool calls.

Advanced. Multi-agent orchestration design

Template + rubric + example

Template

Title: Multi-agent orchestration design

1) Agent roster
| Agent | Role | Capabilities | Limitations |
|-------|------|--------------|-------------|
|       |      |              |             |

2) Coordination pattern
- [ ] Hierarchical (supervisor delegates)
- [ ] Peer-to-peer (agents negotiate)
- [ ] Pipeline (sequential handoffs)
- [ ] Other: 

3) Communication protocol
- Message format: 
- State sharing: 
- Conflict resolution: 

4) Failure handling
- Agent failure recovery: 
- Deadlock prevention: 
- Graceful degradation: 

5) Observability
- Tracing across agents: 
- Performance metrics: 
- Debugging interfaces: 

6) Evidence to keep
-

Rubric (what “good enough” looks like)

Agent roles are distinct and non-overlapping.
Coordination pattern matches the problem structure.
Evidence is identified (what you would log/measure/keep).
Next actions are concrete and realistic.

Worked example (short)

Example (short):
Pattern: hierarchical with planner agent delegating to specialist agents.
Failure handling: if specialist fails twice, escalate to human supervisor.

Capstone. Production agent deployment checklist

Template + rubric + example

Template

Title: Production agent deployment checklist

[ ] Security review completed
[ ] Guardrails tested with adversarial inputs
[ ] Monitoring and alerting configured
[ ] Rollback procedure documented
[ ] User documentation and limitations published
[ ] Incident response playbook prepared
[ ] Cost monitoring and limits set
[ ] Compliance review (if applicable)

Deployment notes:
- 

Known limitations:
- 

Evidence to keep:
-

Rubric (what “good enough” looks like)

All checklist items are addressed or explicitly deferred.
Known limitations are documented publicly.
Evidence is identified (what you would log/measure/keep).
Next actions are concrete and realistic.

Worked example (short)

Example (short):
Limitation: agent cannot access external websites; all data must be provided.
Rollback: feature flag to disable agent and route to manual workflow.