Task statement 1.6 · Agentic architecture and orchestration
Design task decomposition strategies for complex workflows
Before the research coordinator delegates, it splits the job into steps. Some jobs have steps you can list before reading any source, while others reveal theirs only as findings arrive. The wrong shape yields a report that reads as complete but never looked at part of the topic.
share
The tempting move is to write the full subtask list before any research starts, because a fixed list is easy to review and to budget. For an open-ended topic that list holds only what the coordinator already knew, so whatever the first searches uncover has no subtask to land in. For a large review that comes back uneven, the tempting fix is a larger context window, which fails because the input already fits.
KnowledgeTwo ways to split a job
Prompt chaining is a fixed sequential pipeline: the steps are written before the input is read, and each step processes the previous step's output. Each step has one narrow job, so each call is simpler and more accurate. You can also put a programmatic gate between steps, which is code that checks the intermediate output before the next step runs. The gate keeps the chain fixed rather than making it adaptive. Chaining suits predictable multi-aspect reviews. In the research system, every source the document analysis subagent receives goes through the same steps: extract its claims, match each claim to a supporting quote, then record the author and date.
Dynamic adaptive decomposition generates subtasks from intermediate findings instead. The coordinator starts with a step that maps the territory, picks the next subtasks from what that step found, and adds or reorders subtasks as dependencies appear. This suits open-ended investigation, where the useful subtasks are unknown until some of the work is done. Anthropic's research system works this way: its lead agent reads the results and "decides whether more research is needed".
Interactive · which shape does this job need?
Choose a job to see which shape it needs.
Check each source → prompt chaining. The same three steps apply to every source, so they are written before any source is read.
Research an open topic → dynamic adaptive decomposition. Which sectors matter is known only after the mapping search.
Review a 14-file PR → prompt chaining. One local pass per file, then one cross-file integration pass, all known in advance.
Add tests to a legacy codebase → dynamic adaptive decomposition. The high-impact areas appear only once the structure is mapped.
SkillMapping an open-ended topic first
Asked about AI in the creative industries, a coordinator that plans up front assigns digital art, graphic design and photography. Every subagent succeeds, and the report never mentions music, writing or film. The plan was written before any finding existed, so nothing could add to it. The adaptive version maps first and lets each finding change the plan:
// coordinator plan log, adaptive step 1 map broad search → art, design, photography, music, writing, film step 2 rank music and film have the most sources → first step 3 plan one subtask per sector, music and film first step 4 found music licensing turns on one pending court case plan += "analyse the court case" // not in any up-front list
SkillA review split into passes, and a legacy codebase
One review pass over a 14-file pull request gives detailed feedback on some files and superficial comments on others, misses obvious bugs, and flags a pattern in one file while approving identical code in another. Attention dilution is this failure: one pass carries more than it can attend to evenly. The fix is a prompt chain of fourteen per-file local analysis passes, then a separate cross-file integration pass that examines data flow between files. Keeping only findings that two of three runs share would suppress bugs that are caught intermittently.
"Add comprehensive tests to a legacy codebase" is open-ended, so it gets dynamic adaptive decomposition. First map the structure, such as which modules import which. Then identify high-impact areas: a billing module with no tests and 31 callers comes first. Then create a prioritised plan that adapts as dependencies are discovered. When the billing tests show that every charge passes through an untested currency helper, that helper moves to the top.
- Uneven, contradictory findings from one big pass mean attention dilution: split into per-file passes plus an integration pass.
- A plan written before the first finding covers only what you already knew.
- A gate between chain steps checks output. It does not make the chain adaptive.