Understand the idea, test it actively, then explain it in your own words.
You will leave with
Understand continuation across work turn a broad goal into observable milestones, constraints, and stopping conditions distinguish an initial plan from a plan revised after new evidence
Time
25 min
Before you begin
General AI Literacy unit 09 is useful preparation, but the path begins with a complete system-classification foundation.
Do this now
Read the outcomes and useful terms before the first section.
turn a broad goal into observable milestones, constraints, and stopping conditions
distinguish an initial plan from a plan revised after new evidence
Terms you will use
These definitions prepare you for the reading; you do not need to memorize them.Goal
A desired result expressed clearly enough to guide choices and test whether work should continue, change, or stop.
Plan
A provisional organization of steps, dependencies, checks, and alternatives—not a guarantee about what will happen.
Observation
Information returned by the environment after an action, used to update task state and choose what comes next.
01
Continuation needs a usable goal
An agentic system needs more than a topic. “Help with travel” is too open to establish success. “Find three train options arriving before 18:00 within this budget, but do not purchase” defines an output, constraints, and a boundary on action.
Goals can conflict. Speed, price, accuracy, privacy, and user preference may pull in different directions. A system needs priorities or an escalation rule rather than silently deciding which value matters most.
02
Plans are hypotheses about the route
A plan can decompose the goal into searching, comparing, checking constraints, requesting missing information, and presenting a result. Dependencies matter: availability must be checked before a final option is offered, and identity or payment should not be requested when no purchase is authorized.
A useful plan remains revisable. If a tool fails, data conflicts, or a condition changes, the system should update its task state and choose a permitted alternative. Repeating the same failed action is continuation without progress.
03
Milestones, budgets, and stopping rules make work governable
Milestones let a person see whether the system is moving toward the goal. Time, cost, action-count, and retry budgets limit runaway work. A stopping rule can mark success, request clarification, hand off, or terminate safely.
Judge continuation from the full loop: goal, current state, next permitted action, observation, updated state, and stop test. A long chain of polished messages is not evidence of meaningful progress if the state and constraints do not change.
Interactive model
A governed continuation loop
Selected layerGoal and constraints
Define success, exclusions, priorities, and the authority boundary.
Read the complete text alternative
Goal and constraintsDefine success, exclusions, priorities, and the authority boundary.
Plan and next stepChoose a provisional route and one action permitted in the current state.
Observation and updateRead what happened, record relevant state, and revise rather than assume.
Stop, ask, or continueTest completion, uncertainty, budgets, and escalation conditions before another action.
Misconception to correct
“A capable agent should receive a broad goal and work until it decides it is finished.”
Broad goals hide conflicting priorities and unsafe authority. Define constraints, milestones, budgets, approval points, and stopping conditions before continuation.
Active checks
Decide, then compare the reasoning
No grade, score, or streak. Progress records completed learning objects, not your worth or ability.01
Which request gives the safest usable goal for a travel-planning agent?
02
A search tool times out twice. What is the strongest next-step design?
Explain back and transfer
Turn a request into governed continuation
Choose a multi-step request. Write its result, constraints, first three milestones, one reason to replan, and the conditions to stop, ask, or escalate.
Answer both checks before completing.
Source context
What supports this unit and where it stops
Evidence status: established
Hugging Face Agents Course — IntroductionHugging Face
Offers a practical model of iterative thought, action, and observation for technical learners; implementations differ and internal reasoning need not be visible or reliable.
AI Risk Management FrameworkU.S. National Institute of Standards and Technology
Supports defining context, objectives, affected actors, risk, monitoring, and management around an AI system; it does not prescribe one planning algorithm.
inAi