What is task decomposition in an AI system?

Breaking a complex goal into smaller tasks the AI can handle reliably. An agentic system pursues a goal over several steps rather than answering once, and large goals are rarely safe to give a model all at once, so the work has to be split. How it is split largely decides how reliable the system is.

When should I use a fixed pipeline instead of an adaptive plan?

Use a fixed sequential pipeline when you already know the correct steps: invoice processing, form extraction, compliance checks, standardised document review. Use dynamic adaptive decomposition when the right next step depends on what the system discovers, such as debugging an unfamiliar system or investigating a production incident. A planner rediscovering five steps you already knew costs tokens, latency and run-to-run variation for nothing.

What is attention dilution?

When an AI system processes too many separate items in one pass and gives them uneven attention. In a fourteen-file code review the first few files get detailed feedback and the last few get shallow feedback, while the final report still looks complete. It is a structural problem caused by how the work is arranged, so a better prompt only softens it.

Does a bigger context window fix attention dilution?

No. A larger context window lets you put more text into one call, but it does not guarantee the model treats every part of that text with equal care. Room to include more material and care taken over each part of it are different things. The fix is to give each item its own pass and then add a separate integration pass across the results.

Task Decomposition for AI Agents: Fixed Pipelines, Adaptive Plans, and the Multi-Pass Fix

Large goals are rarely safe to give an AI model all at once, so the work has to be broken into pieces. This project builds a multi-pass code review pipeline from scratch to show why: a single-pass baseline that reviews fourteen files in one prompt and quietly stops reading, a per-file layer that gives the fourteenth file the same attention as the first, and an integration pass that catches what no per-file review can see.

A teacher has 14 exam papers to mark.

The first few papers get careful notes: corrections in the margin, comments on weak arguments, suggestions for improvement. By paper 10, the feedback is much thinner. By paper 14, the teacher may still assign a grade, but the comments are brief and some obvious mistakes slip through.

The teacher did not suddenly forget how to teach. The problem is the way the work was arranged. Too much material was handled in one long pass, so the quality became uneven.

The same thing happens in AI systems.

An agentic AI system is an AI system that does more than answer one prompt. It can pursue a goal through several steps, often using tools, meaning external abilities such as reading files, searching documents, running code, querying a database, or calling an API. A simple chatbot answers once. An agentic system can plan, act, observe the result, and continue.

That extra power creates a design problem: large goals are rarely safe to give to the model all at once. The work has to be broken into pieces. That is task decomposition, and there are three shapes it usually takes.

Fig. 1 · Three shapes of the same job. The same goal, arranged three ways, on one shared layout so the shapes can actually be compared. A fixed pipeline knows its route before it starts. An adaptive plan revises itself when the evidence moves, returning to an earlier step rather than abandoning the plan. A hierarchical plan has phases that are each decomposed again. Choosing between them is the first real design decision in an agentic system.

Choosing badly is expensive in a direction that is easy to miss. Good decomposition makes the agent more reliable. Bad decomposition makes it look clever at first and then fail in quiet, inconsistent ways.

Sign in to view this build.

Create a free account to access the full build walkthrough on kevinsomany.com.

No account yet? Sign up free and come back here.

Task Decomposition for AI Agents: Fixed Pipelines, Adaptive Plans, and the Multi-Pass Fix

Sign in to view this build.

Sign in or create an account.

Check your inbox.

Something went wrong.