orchestration ยท multi-model mixes & agent frameworks

When does wiring models together actually help?

A single model is one agent in a loop. A mix wires several over one workspace - a planner, an implementer, a critic, a judge. lemon-squeezer runs the first four of these as scored harnesses on the same evals as everything else; the fifth, Swarm, is shown for shape (it is how this very page was built, not a benchmark arm). Below: how each is wired, and what it buys.

Planner
Implementer
Critic
Judge
Tester

Architect

A strong model plans; a cheap one writes the code.

A strong model plans; a cheap one writes the code.workspaceplanplannerPlannerimplementerImplementer

Critique

Draft, get reviewed, refine - on a loop.

Draft, get reviewed, refine - on a loop.workspacedraftrefineimplementerImplementercriticCritic

Ensemble

Several models each draft; a judge picks the best.

Several models each draft; a judge picks the best.workspacepickimpl AImplementerimpl BImplementerimpl CImplementerjudgeJudge

Verify

Write code and its tests, then self-correct until they pass.

Write code and its tests, then self-correct until they pass.workspacetestfiximplementerImplementertesterTester

Swarm

Many agents build disjoint parts in parallel; a conductor integrates.

Many agents build disjoint parts in parallel; a conductor integrates.workspaceintegrateagent AImplementeragent BImplementeragent CImplementerconductorJudge

Do mixes beat the best single model?

Loading runs...

Built by orchestration

Claude Code conductor ยท claude-opus-4.8

planned the work, wrote the page, integrated the pieces, verified the build

codex

gpt-5.5 โ†—cloud

the animated orchestration flow diagram (278 lines)

ChatGPT sub

kimi (via squeezer)

kimi-k2.7-code โ†—cloud

the orchestration patterns data (JSON)

$0.011 ยท 23k/2.7k tok

aider

deepseek-v4-flash โ†—cloud

this conductor panel

$0.0002

Three agents in parallel: ~1.1c of metered API for the two paid legs (kimi + aider); codex ran on a ChatGPT subscription and Claude was the conductor. Run concurrently, so wall-clock was the slowest single leg, not the sum.

This tab was built by one Claude conducting codex, kimi, and aider in parallel - the same idea as the mixes above, one level up. Writeup: nanlives.com.

References & frameworks