1. Design the state machine
Outcome: Agent behavior does not depend on a hidden loop.
Tasks
- →Define states
- →Define transitions
- →Add maximum-step limits
- →Add terminal outcomes
Checks
- ✓Every state has an exit condition
- ✓An infinite loop is impossible
Advanced•16–24 hours
Governed Operations Agent
Build a governed AI agent with tools, explicit state, approvals, cost and time limits, an audit log, and a rollback path.
agent statetool contractsMCPauthorizationhuman approvalauditability
Scenario
Task
An agent prepares changes to an operating environment: it reads data, creates a plan, invokes tools, and requests approval before risky actions. Every action must be reconstructable from the audit trail.
Step-by-step execution
2. Introduce tool governance
Outcome: Every tool has a contract and explicit access boundaries.
Tasks
- →Define input/output schemas
- →Assign a risk level
- →Set the permission scope
- →Add dry-run support
Checks
- ✓Unknown tools are blocked
- ✓Write tools never run without an applicable policy
3. Add approvals and budgets
Outcome: Risk, time, and cost are bounded before execution.
Tasks
- →Define approval gates
- →Add token/cost budgets
- →Add a timeout
- →Add cancellation
Checks
- ✓High-risk actions always wait for approval
- ✓Budget exhaustion ends the cycle safely
4. Verify auditability and rollback
Outcome: An action can be explained, replayed, or rolled back.
Tasks
- →Log decision inputs
- →Persist tool results
- →Add a rollback handler
- →Run a failure drill
Checks
- ✓The audit log reconstructs the sequence
- ✓Rollback has evidence
Acceptance criteria
- A deterministic state machine exists
- Every tool has a risk and permission policy
- Write actions require approval
- The agent has time/cost/step limits
- The failure drill and rollback pass
Assessment rubric
How the result is assessed
Passing score: 75/100 · Distinction: 92/100
State machine and stop conditions
States, transitions, terminal outcomes, and recovery are explicit.
Insufficient
Behavior depends on a hidden or unbounded loop.
Competent
Every state has an exit condition and explicit limits.
Strong
Deterministic replay, recovery states, and property tests are implemented.
Evidence required
- ✓ Link to code or artifact
- ✓ Short README explaining decisions
- ✓ Test output or runtime evidence
- ✓ State diagram or transition table
Tool governance
Every tool has a schema, permission scope, risk level, and dry-run policy.
Insufficient
Tools are invoked without explicit contracts or access boundaries.
Competent
The registry defines contracts, risks, and permissions.
Strong
Policies are enforced by a separate policy engine and covered by tests.
Evidence required
- ✓ Link to code or artifact
- ✓ Short README explaining decisions
- ✓ Test output or runtime evidence
- ✓ Tool registry
Approvals and budgets
Risky actions, time, steps, and cost are bounded before execution.
Insufficient
Write actions can run without approval or limits.
Competent
Approval gates and budgets block unsafe actions.
Strong
Risk-based policy, cancellation, and delegated approvals are implemented.
Evidence required
- ✓ Link to code or artifact
- ✓ Short README explaining decisions
- ✓ Test output or runtime evidence
- ✓ Example approval request and budget-exhaustion result
Auditability and rollback
The decision sequence is reproducible and changes have rollback evidence.
Insufficient
The audit trail is incomplete or rollback is only declarative.
Competent
The audit log reconstructs the cycle and rollback is verified.
Strong
Tamper-evident logging, replay, and an automated recovery drill are implemented.
Evidence required
- ✓ Link to code or artifact
- ✓ Short README explaining decisions
- ✓ Test output or runtime evidence
- ✓ Failure drill and rollback report