Planning and Task Decomposition for Agents
Teach agents to break big goals into checkable steps, revise the plan and know when they are done
A taste of a lesson
My research agent writes a report after reading two sources, even when I ask for a thorough review. How do I stop that?
That is premature completion, and 'thorough' is too vague to act on. Give it a definition of done instead: at least six sources, at least two that disagree, every claim cited, and a short list of open questions. Ask for a plan first that lists which kinds of sources it will look for, then have a final step that checks the draft against each criterion and goes back if one fails. Try writing your own acceptance criteria now: what would make you call the report finished?
Written by the teacher as an example. In your lesson the tutor answers your own questions, and like any AI it can be wrong.
What you will be able to do
- Recognise drift, premature completion and thrashing in agent runs
- Choose between implicit planning, plan then act, and a separate planner step
- Write steps that each produce a checkable output
- Define acceptance criteria so an agent knows when it is done
- Measure whether explicit planning improves results on your tasks
Lesson plan
- 1 How long tasks fail Name the three common failure patterns of multi step agent tasks. Start
- 2 Four ways to plan Compare implicit planning, plan then act, plan act revise, and a separate planner. Start
- 3 Steps you can check Decompose a goal into steps that each produce a verifiable result. Start
- 4 Defining done Write acceptance criteria that stop both early quitting and endless polishing. Start
- 5 Revising the plan Let the agent update its plan when new information arrives without losing the goal. Start
- 6 Measuring whether planning helps Test explicit planning against simpler designs on your own tasks. Start
Try asking
About this tutor
For developers and technical product people whose agents handle tasks with several stages: research then write, investigate then fix, gather then report. Big goals fail when an agent charges ahead without a plan or clings to a plan that stopped making sense. You will learn when to ask the model for an explicit plan, how to make each step verifiable, how to let the agent revise the plan after new information, and how to define done. We compare planning inside one agent with a separate planner step in a workflow, and look at the costs of each.
Reviews
4.5
2 ratingsSample
- Arjun S.Sample
The lesson on measuring whether planning helps was the surprise. For our short tasks it did not, so we removed it and saved money.
- Chloe D.Sample
The three failure patterns gave me words for what I was seeing. Acceptance criteria fixed most of my early stopping problems. Wish there had been more on long coding tasks.
About the teacher
I teach how AI agents are built: the loop, the tools, the memory, and when a plain workflow is the better choice
9 tutors 310 lessons taught Sample
I build and teach the inner workings of AI agents. Most of my working life has been spent on backend systems, so I approach agents the way I approach any distributed system: what runs, in what order, what can fail, and what it costs. I like to start every topic with a drawing of the loop on a whiteboard and...
See Hiroshi's profile and tutorsMore like this
Other tutors on the same or nearby topics.