Skip to content
anythingai.space
Agentic Systemsintermediate12 min · 9 steps

Planning and Decomposition

Turning a goal into a plan an agent can follow: clarify first, choose static or dynamic planning, decompose recursively within limits, and pay for revision.

few answersbarely readHYPER-TAU-BENCH · HELD-OUT TESTS PASSEDAgent alone23.9%Agent + engineer82.2%Client20–25 requirementsPlannerreasoning model≤ 4 questionsBusiness records< 80 of ~1,700 opened
requestresponsedataerrorsignal
1 / 9
ProblemStep 1 of 9

Agents decompose the wrong goal — and rarely ask

In Sierra's hyper-tau-bench (September 2026) a developer agent had to recover requirements from business records and a client interview, then build a customer-service agent. Claude Opus 5 working alone passed 23.9% of held-out tests; paired with an engineer who had deep context, 82.2%.

The agents opened fewer than 80 of ~1,700 available files and asked at most 4 client questions — when the client alone held 20 to 25 requirements. OSWorld 2.0 names the same gap for desktop agents: missing information, because agents rarely ask.

Watch outThe most expensive planning error happens before the first step: a perfect plan for the wrong goal.

Lesson 6 of 15 in Agentic Systems