LLM Council

Instead of paying for one frontier model, convene a council of cheaper models that deliberate and a judge that synthesizes — and you can rebuild near-frontier intelligence on demand. The reframe: the winner won't have the smartest model, they'll have the best loop. Source: X/@EXM7777, 2026-06-12; X/@AnatoliKopadze, 2026-06-20

The thesis

When a frontier model gets capped or pulled (the post's framing: "Fable 5 felt like AGI, then it was killed; Opus 4.8 feels lobotomized by comparison"), you don't have to accept the downgrade. You can reconstruct the intelligence from cheaper models by running them as a council: several models answer in parallel, then a judge/aggregator reconciles into one answer that beats any single member. Run it on demand, pay per use, swap members freely. Source: X/@EXM7777, 2026-06-12

This is the same mechanic Kevin already tracks as OpenRouter Fusion (a productized council — panel + judge exposed as one model alias) and as a topology in Multi-Agent Topologies (the debate and evaluator-optimizer patterns). The council is the ensemble expression; the loop below is the iterative expression.

Council + loop are the same family

@AnatoliKopadze's writeup of the Anthropic/Claude-Code team's build (plan → build → judge agents cycling until the app works) makes the load-bearing point: "the winners won't have the smartest model, they'll have the best loop." A council deliberating once and a loop iterating until a judge passes are points on one spectrum — both spend more tokens across more model-passes to buy quality a single forward pass can't. The reviewed video clip is an AI Engineer Europe talk titled "How to Build Agents That Run for Hours (Without Losing the Plot)," and Anthropic's official writeup maps the pattern to a planner, generator, and evaluator with Playwright-based review. See Browser Testing Skills (the build loop), The Eval Loop (Slop Is an Output Problem) (the judge/gate), and Borrowed Intelligence (Drain It into Plans) (drain strong-model planning, execute cheap). Source: X/@AnatoliKopadze and local video review, 2026-07-04; Source: Anthropic, 2026-03-24

How to build one (the council recipe)

• Members: a few diverse, cheap-but-strong models — e.g. GLM-5.2 (open-weights, long-horizon coding) alongside other open/hosted models via a gateway. Diversity of training > redundancy.
• Protocol: parallel independent answers → optional critique/debate round → a judge synthesizes (or votes). This is Multi-Agent Topologies's evaluator-optimizer / debate topology.
• The economics gate (from Multi-Agent Topologies): a council spends multiples of single-model tokens, so it's justified only for high-value, parallelizable, quality-sensitive work — deep research, architecture critique, hard one-shots — not routine coding.
• Shortcut: OpenRouter Fusion already packages this as openrouter/fusion; use it before hand-rolling a council.

The reviewed setup diagram adds the policy layer this page previously implied but did not spell out. A council is not a router: a router switches models, while a council runs several models in parallel and fuses their outputs through a judge/frontier model. Convene it for deep research, codebase migration, long orchestration jobs, briefing a tighter agent, or building a knowledge base; skip it when a cheap model can answer directly. Source: X/@EXM7777 and local image review, 2026-07-04

Tiny Coordinator Watch: TRINITY

The TRINITY bookmark is a research-watch item for this page's central claim: a small coordinator can sometimes buy more quality by routing, fusing, and managing model work than by answering directly. The linked paper is arXiv 2512.04695; the local artifact is an annotated paper screenshot rather than a runnable system. Treat it as evidence that the coordinator/council direction is active research, not as a production architecture recommendation until the tasks, baselines, and reproducibility details are reviewed. Source: X bookmark, 2026-07-02; Source: arXiv, 2026-07-06

For Kevin, the actionable takeaway is not "use a tiny model as the answerer." It is to keep coordinator roles explicit: planner, router, judge, synthesizer, and budget controller are distinct jobs. If a cheaper coordinator can select or fuse better workers, that belongs in Orchestration Variables and The Eval Loop (Slop Is an Output Problem) as a measurable harness variable. Source: X bookmark artifact audit, 2026-07-06

Role protocol

The OpenRouter MCP artifact shows the practical protocol: use a small council with differentiated jobs and a final editor, not a majority vote or a 12-model pileup. The reviewed screenshot proposed five roles: chair/final spec editor, implementation architect, fullstack red-team implementer, database/systems skeptic, and cost/pragmatism reviewer. The point is to assign failure modes to members, then let an editor reconcile tradeoffs into a spec or decision. Source: X/@lennox_saint, 2026-06-26; Source: local screenshot review, 2026-07-03

OpenRouter's own example makes the role protocol situational: tune the council to the codebase and use case. This is where OpenRouter MCP fits: it gives the coding agent live model/pricing/docs lookup and billable chat-send test calls while the council is being designed. Use that to pick members and run small prompt trials; keep durable production inference on the OpenRouter API or the app's chosen gateway. Source: X/@OpenRouter, 2026-06-26; Source: OpenRouter MCP docs, 2026-07-04

Single-task second opinions

Theo's Codex loop is the smallest council worth keeping: after Codex designs an API, ask another frontier model for a second opinion before the implementation hardens around the wrong interface. In a local harness, that can be as simple as claude -p over the API sketch. The pattern is not "always ask another model"; it is "pay for a second model at irreversible boundaries." Source: X/@theo, 2026-06-21

Programmable routing gateways

OrcaRouter is the infrastructure-shaped version of the same thesis. Instead of only exposing a static "fusion" slug, it gives operators a routing DSL over request shape, task class, headers, tenant context, and agent session state. A rule can route easy tasks cheaply, escalate hard code, fan out to a panel, or use an arbiter such as best_of_n, synthesize, or tests_pass. Source: OrcaRouter docs, 2026-06-30

This turns the council into a versioned routing policy. The quality primitive is no longer just "which model?" It is "which graph fires under which evidence, with what rollout, audit trail, and rollback path?"

Relevance to Kevin

Direct fit for Dedalus / agent infra: the council/loop is how you get frontier-grade output from a fleet of cheaper models you control — resilient to any one lab capping a model. It's the orchestration-variable view (quality is a knob you tune with passes + judges, not just model choice) and reinforces the Multi-Agent Topologies "fleet, not chat" framing. Build vs buy: openrouter-fusion to buy it, the council recipe to own it.

The hard-SaaS caveat: councils and loops buy more attempts, more perspectives, and more verification budget. They do not eliminate product time. Kevin's Dedalus Machines launch discussion treats LLM councils, long-reasoning models, and agent loops as harnesses around bare-model limits, not proof that software can be one-shot into a bug-free maintained product. Source: raw/notes/saas-bug-moat-slack-thread-2026-06-22.md

---

Timeline

• 2026-07-06 | Added TRINITY as a research-watch item for tiny coordinators and model councils. Local decision: record the arXiv signal and local visual artifact, but require task-matched eval/repro review before treating it as an operational routing design. Source: X bookmark, 2026-07-02; Source: arXiv 2512.04695
• 2026-07-04 | Added Theo's claude -p second-opinion loop as the smallest council pattern: ask another frontier model at irreversible design boundaries such as API shape, then continue implementation with the reconciled judgment. Source: X/@theo, 2026-06-21
• 2026-06-22 | Captured the LLM-council thesis (EXM7777: rebuild Fable-5-level intelligence from a council of cheaper models) and the loop reframe (AnatoliKopadze: "best loop wins, not smartest model"). Wired to OpenRouter Fusion, GLM-5.2, Multi-Agent Topologies, and the loop family. Source: X/@EXM7777, 2026-06-12; X/@AnatoliKopadze, 2026-06-20
• 2026-06-30 | Added OrcaRouter as the programmable-routing version of the LLM-council thesis: DSL rules, agent-state routing, fan-out, arbiters, shadow/canary rollout, and rollback. Source: OrcaRouter docs, 2026-06-30
• 2026-06-22 | Added the hard-SaaS caveat: model councils are quality-amplifying harnesses, not a replacement for real product usage, regression, and maintenance. Source: raw/notes/saas-bug-moat-slack-thread-2026-06-22.md
• 2026-07-03 | Added the OpenRouter MCP model-council artifact: small role-diverse council, final editor, and live model/cost lookup inside the coding agent. Source: X/@lennox_saint, 2026-06-26
• 2026-07-04 | Added OpenRouter's first-party MCP model-council example: tune the council to the current codebase/use case and use live model/cost/docs lookup while designing the roles. Source: X/@OpenRouter, 2026-06-26
• 2026-07-04 | Deepened the Anthropic app-builder loop with local video review and the official Anthropic harness-design post: plan/build/judge is the iterative sibling of a model council, using more passes and an evaluator to buy quality. Source: X/@AnatoliKopadze, 2026-06-20; Source: Anthropic, 2026-03-24
• 2026-07-04 | Added the EXM setup diagram as the council policy artifact: a router swaps models; a council fuses multiple model outputs through a judge and should be convened only when the task is high-value enough to pay for parallel reasoning. Source: X/@EXM7777, 2026-06-19