Control structures for agent work
One runtime / 32 control structures / agents you own
We buildthe structure.You keep the work.
Better agent work does not come from one more perfect prompt. It comes from deliberate generation, pressure, reduction, agreement, and a contract that knows when to stop.
RUN SPECIMENCHALLENGE–BEAM / N3
scored challenged reducer artifact
01
Why structure
A single ask is a gamble.
Every team building with agents finds the same ceiling: quality plateaus at one model's one attempt. The remedies are already known. Make many and prune. Expose work to adversarial pressure. Reduce in stages. Require agreement. Stop when stable.
02
A run, reduced
Not a chat log / not a black box
Work moving through a system.
Five agents and two system lanes move across three phases. Lines branch, thicken, loop, and converge. Hollow events are ordinary work; blue is scored, coral is conflict, diamonds are milestones, and solid events leave artifacts.
05agents
39events
03phases
01git-backed run
03
The pattern library
Known structures.
Executable contracts.
Faithful, cited reproductions form the core. Every pattern speaks the same phase contract, so a winner can face challenge, revision, and compression in one composed run.
Beam loop
Generate candidates. Score them as peers. Carry only the strongest work forward.
Beam search — Lowerre (1976); Yao et al. (2023), Tree of ThoughtsExpanding elaboration
Start with a sketch. Expand its parts in parallel. Rejoin them with added depth.
Ning et al. (2023), Skeleton-of-ThoughtEvaluator loop
Produce, evaluate, and repeat until the work crosses a declared threshold.
Madaan et al. (2023), Self-RefineDialectic
Put two positions under pressure and turn their conflict into a stronger synthesis.
Du et al. (2023), multi-agent debate; Irving et al. (2018)04
Three commands