Give three AI agents the same instruction, "research this market," and you will often get the same answer three times, at three times the cost.
The fix Anthropic landed on was not smarter agents. It was better tickets.
Agent planning is the step where the model stops and works out what needs doing, and in what order, before it calls a single tool (a tool being anything it acts with: a web search, a database query, a file write). Task decomposition is the splitting: one fuzzy goal becomes subtasks with clear finish lines. "Research this market" becomes "find competitor pricing" and "summarize recent funding rounds."
The difference lives in what gets handed over. Anthropic's lead agent gives each worker four things: the objective, the output format it wants back, which tools and sources to prefer, and boundaries so two workers do not chase the same thread. Pass along a bare "research topic X" instead and they duplicate work and misread the assignment.
That planning step carries most of the gain: a lead-plus-workers setup beat one strong agent by 90% on Anthropic's internal research tasks, with the workers running the smaller, cheaper model.
It is not free. Agents burn roughly 4x the tokens (the chunks of text a model is billed on) of a normal chat turn, and multi-agent setups roughly 15x. Splitting only pays when the pieces are truly independent. If every worker needs the same context, or step two waits on step one, one agent is cheaper.
Which makes the real skill scoping, not prompting.
Quick check before you scroll: Why does Anthropic's lead agent write out detailed instructions (objective, output format, tools, boundaries) for each subagent instead of just forwarding the raw user query to several agents at once?
Full breakdown + the answer: frankduah.me/learnings/2026-09-02-agent-planning-and-task-decomposition
New here? I post a bite-size AI / ML concept like this every day - follow me for the daily drop, and it compounds fast. Why I do it: https://lnkd.in/gK8knHDH
#AgentPlanning #AI #LLM #AIAgents #MachineLearning
The answer
Without clear scope, parallel agents tend to duplicate each other's work or leave gaps - explicit task boundaries per subagent are what make decomposition actually pay off.