The architectures
Six ways to put structure around a model.
Each is the same Claude model with the same tools and the same task. Only the harness around the model changes: how many calls it makes, what each call is shown, and who gets to say no.
Single
Is any structure worth its price at all?
One agent, one context, the full toolset. The control condition every other architecture has to beat.
1 stagefull toolset
Plan-execute
Does committing to a sequence up front beat deciding one action at a time?
A tool-less planner writes the steps; a second call carries them out with the full toolset.
2 stagesno toolsfull toolset
Orchestrator-workers
Does splitting a task across parallel contexts pay for the coordination?
An orchestrator splits the task, workers run the parts in parallel contexts, and a final call merges their reports.
3 stagesparallelno toolsfull toolset
Plan-execute-verify
Is a second opinion on the finished state worth the cost of asking for it?
Plan and execute, then a verifier with read-only tools checks the real state and can send the work back.
3 stagesno toolsfull toolsetread-only tools
Propose-select
Does choosing among independent proposals beat committing to the first idea?
Three proposers investigate in parallel with read-only tools, a selector picks one plan, and an executor applies it.
3 stagesparallelread-only toolsno toolsfull toolset
Gated-execute
Is review at the point of action worth a model call per action?
One agent acts, but every state-changing tool call must pass a read-only reviewer before it runs.
2 stagesfull toolsetread-only tools
How to read a glyph
Every architecture is drawn the same way on every page of this site: each call is a shape, and a line runs from the call that hands work on to the call that receives it. A shape means one thing in all six.
- Agent
- AgentA model call that acts, with tools.
- Worker
- WorkerA smaller call in a parallel pool, or a state change.
- Reasoner
- ReasonerA call with no tools: it can only think and reply.
- Decision
- DecisionA verifier or selector: it chooses and may send work back.
- Gate
- GateEvery action passes through it before it runs.
- Reviewer
- ReviewerA read-only call standing beside a gate.
Provenance of the numbers on this page
- run set
- scaffold-or-scale-v1
- models
- Haiku 4.5, Sonnet 5.5, Opus 5.5
- effort
- high
- repeats
- 2
- fixtures
- replay
- runs
- 900 / 900
- strategy cost
- $354.44
- judge cost
- $9.19
- wall clock
- 9h 48m
- git
- fce5807b2d
Cost excludes judge calls, which are tracked separately so a scoring model never flatters or penalises an architecture. Every figure on this site is read from results/ at build time.