PrerequisiteAI 104 recommended, plus Modules 1 to 5 of this course
0 / 6 completed
Sync my progressJoin a session to keep progress across devices and share it with your instructor
Learning objectives
Describe the adaptive planning model: Plan, Act, Observe, State Changed?, then Continue, Re-plan, Escalate or Stop
Recognize eight re-planning triggers and explain what each one signals
Distinguish plan monitoring, plan invalidation, goal revision and a recovery decision
Define bounded autonomy with allowed and prohibited actions, budget, limits, escalation, approval, stop and human override
Complete Activity 6 by deciding whether to continue, re-plan, escalate or stop, and explaining why
Adapting a plan
10 min
When the world changes, the plan has to change too
A plan is made with what the agent knew at the time. Once it acts, the state can change: new information arrives, a step fails, a constraint tightens. An adaptive agent checks after each action whether its plan still fits the world.
Core question
When should an agent stop following the original plan and plan again?
Adaptive planning model
Plan: the current plan is in place
Act: the agent takes the next step
Observe: the agent looks at the result and the state
State Changed?: if no, Continue with the plan
If yes: Re-evaluate the plan
Then choose one: Continue, Re-plan, Escalate or Stop
Info
Text version of the diagram: Plan leads to Act, then Observe, then a question, State Changed? A no means Continue. A yes means the agent re-evaluates the plan and then chooses one of four outcomes: Continue, Re-plan, Escalate or Stop.
Info
Human, AI, Human: a person defines the task and its limits, the agent adapts within them, and a person reviews and remains responsible when the situation goes beyond those limits.
Key takeaway: After every action, observe, ask whether the state changed, and choose Continue, Re-plan, Escalate or Stop.
10 min
Re-planning triggers
A trigger is a condition that makes the current plan questionable. Naming triggers in advance turns a vague feeling that something is off into a clear rule.
Eight common re-planning triggers (simulated examples)
Trigger
What it signals
Simulated example
Required information unavailable
A step cannot run as planned
The report the plan depends on has not been published yet
Tool unavailable
An action in the plan cannot be taken
The calendar service does not respond
New evidence contradicts assumptions
The plan rests on something that is no longer true
A newer policy replaces the one the plan followed
Plan exceeds cost limit
The route is too expensive
The planned searches would use more than the budget
Risk increases
The next step is now more dangerous
The task turns out to involve personal data
User changes objective
The goal itself has moved
The requester now wants a comparison, not a summary
Confidence drops
The agent is less sure of its path
Two sources disagree and neither is clearly better
Human rejects intermediate result
A person says this is not what they wanted
A reviewer rejects the draft outline
Warning
A trigger does not always mean Re-plan. Some triggers call for Escalate or Stop instead, for example rising risk or a person rejecting the result.
Check yourself
Midway through a plan, the requester says they now want a different deliverable. Which trigger is this, and what is the natural response?
Key takeaway: Name your triggers in advance, and remember that a trigger may call for Escalate or Stop, not only Re-plan.
10 min
Monitoring, invalidation, goal revision and recovery decisions
Adapting well takes four separate ideas. Keeping them apart helps you decide what to do when something changes.
Four ideas in adaptive planning
Idea
Question it answers
Example (simulated)
Plan monitoring
Is the plan still on track after this step?
Check each result against what the plan expected
Plan invalidation
Has something made part or all of the plan no longer valid?
A step depends on a document that was withdrawn
Goal revision
Has the goal itself changed or become unreachable?
The requester narrows the scope to one department
Recovery decision
What do we do now: Continue, Re-plan, Escalate or Stop?
Choose to re-plan the remaining steps only
Re-plan
The goal still stands
The route to it needs to change
Example: switch to another source of evidence
Revise the goal
The goal itself has moved or cannot be reached
A person usually confirms the new goal
Example: the deliverable is redefined with the requester
Info
Boundary of this course: this module covers the decision to adapt. How to save progress, resume after a crash or build recovery infrastructure belongs to a future advanced course and is not taught here.
Check yourself
A plan step relies on a document that has just been withdrawn, but the goal has not changed. What is the best reading?
Key takeaway: Monitor the plan, notice invalidation, revise the goal only when it has moved, and then make a clear recovery decision.
Bounded autonomy
10 min
Bounded autonomy: how far may the agent adapt?
An agent that can change its own plan needs limits. Bounded autonomy means the agent adapts freely inside a boundary that a person has defined, and hands over when it reaches the edge.
What every adaptive agent should define
Element
What it sets
Example (simulated)
Allowed actions
What the agent may do on its own
Read approved documents, draft a summary
Prohibited actions
What the agent must never do
Send anything outside the organization
Budget
How much cost the agent may spend
At most a set number of searches
Time / step limit
How long or how many steps it may run
No more than a set number of re-plans
Escalation condition
When it must hand the situation to a person
Confidence stays low after one re-plan
Human approval condition
Which actions need an explicit yes first
Any action that commits the organization
Stop condition
When it must halt
Budget used up, or risk becomes high
One more safeguard sits above all of these: human override. A person can pause, redirect or cancel the agent at any time, and the design should make that easy.
Warning
Without a re-plan limit, an agent can keep re-planning forever. Set a maximum, and treat reaching it as a reason to escalate or stop.
Check yourself
An agent has re-planned several times and is still not making progress. What should a well-bounded design do?
Key takeaway: Define the boundary first: allowed, prohibited, budget, limits, escalation, approval, stop, and human override.
Practice and next steps
8 min
Activity 6: Re-plan or Continue?
You will be given a scenario in which the situation changes partway through a plan. For each scenario, decide what the agent should do.
1Inspect the new state What changed, and how do you know?
2Check the plan Does the original plan remain valid, fully or partly?
3Choose Continue, Re-plan, Escalate or Stop.
4Explain why Name the trigger and the boundary that supports your choice.
The four choices
Choice
Choose it when
Continue
The change is small and the plan is still valid
Re-plan
The goal stands but part of the route no longer works
Escalate
The situation is high-impact, unclear or outside the allowed boundary
Stop
A stop condition is met, or continuing is unsafe or not worth the cost
Warning
Privacy reminder: use simulated or already-public scenarios. Do not enter company secrets, client personal data or unauthorized sensitive information.
Midway through a task, new information shows the next step would involve personal data the agent is not allowed to handle. What is the best choice?
Key takeaway: Inspect the new state, check the plan, choose one of four outcomes, and explain it.
7 min
Your Re-planning Policy and next steps
Done
Deliverable of this module: a Re-planning Policy. It states which triggers you watch, what the agent does for each (Continue, Re-plan, Escalate or Stop), and the bounds inside which it may adapt.
Triggers: which of the eight, or your own, apply to your task
Response: Continue, Re-plan, Escalate or Stop for each trigger
Bounds: allowed and prohibited actions, budget, limits, approval and stop conditions
Human role: when a person is asked, and how a person can override
In Module 7 you will place this policy on the Agent Planning & Reasoning Canvas, in the Re-planning Triggers, Constraints & Boundaries, Human Escalation and Stop Conditions sections.
