AIneedsasheepdog.
Three systems that route every task to the right model, make models argue over the hard calls, and stop runaway sessions before they burn. Built in public, with receipts.
One task, three jobs, one trace.
mechanism: junction: three candidate tracks, no commitment yet
A task arrives. It could go to any model: the expensive genius, the fast intern, the cheap workhorse.
Switchboard weighs quality, cost, and latency, then commits with a written reason.
Some questions are too important for one opinion. Quorum convenes the classroom.
While work runs, Governor watches, and it knows the difference between a long job and a runaway one.
Output, verified. With the receipt.
Every decision above is real behavior from real systems. Scroll on to operate them.
// what you can do with these
Operations
evolved
- Go to SwitchboardExplainable RoutingCost GuardrailsCross-Lab AuditJSONL Traces
Switchboard
Routes every request to the cheapest model that is qualified for it, and writes down why.
--% cheaper
waiting for live catalog
● LIVERead the code → - Go to QuorumNumbered ClaimsBlind CritiqueMinority ReportArbiter Verdict
Quorum
Makes three models argue a hard question and preserves the dissent, not just the verdict.
● BUILT · unmeasuredRead the code → - Go to GovernorFive-Rung LadderNo LLM InsideHysteresisHuman Authority
Governor
Watches running agents and tells a runaway loop from a healthy long day of work.
● RELEASED v0.2.0Read the code →
Switchboard
Every AI product decides which model gets each task, thousands of times a day. Most teams decide once, by hardcoding a model name, and pay for it invisibly ever after.
policy.max_provider_failovers = 2 · outages, rate limits and timeouts reroute to the runner-up · and yes, Fable and Luna are real models in the live catalog below
A dispatcher for AI work.
Switchboard decides before every call which model gets the job, writes down why, and keeps the receipt.
- 01
Gates before scores.
Hard work goes frontier-only, and no model wins a task it is not rated for, however cheap. When nothing qualifies it escalates upward and says so.
- 02
Chooses with a written reason.
Quality, cost and latency, weighed by a policy you can read in a diff. Every decision returns a ranked list and a plain sentence you can argue with.
- 03
Verifies and logs.
A model from a different lab audits the output, so shared blind spots do not grade their own homework. Everything lands in a JSONL trace you can replay.
loading live catalog…
fetching catalog from GitHub...
Not the same job as Anthropic's Advisor tool
Advisor tool facts as of its 2026 beta docs.
A cheaper Claude consults a smarter Claude mid-task. The model decides when to ask for help.
You decide before the task which model gets the job, by a policy you own and version.
One vendor's family: Claude models only, inside one API request.
Routes across labs. The starter catalog spans Anthropic, OpenAI, Google and DeepSeek, and audits prefer a different lab than the one that did the work.
Guidance happens inside the request, invisible to your policy layer.
Every choice is a logged, replayable decision with a written rationale and a cost attached.
They compose. Switchboard picks the executor and writes down why. The executor can still ask an advisor for help mid-task.
Quorum
For decisions that matter, eloquence is not evidence. Quorum makes three models argue and shows you the fight.
convening rule: 9.3% of a 43-task workload convenes · 4/4 genuine judgement calls convene · 0 routine false positives
A deliberation protocol.
Three models answer in isolation, attack each other's claims by number, revise, and an arbiter delivers the verdict with the dissent preserved.
The gates fail closed, the opposite of Switchboard, on purpose: a wrong convene costs a 7x bill, so the default answer is no.
- 01
Isolates first.
Three answer sheets are written blind, then anonymized to A and B. Style never leaves the room.
- 02
Attacks claims by number.
Critique targets claim 3, not vibes. Vague agreement is structurally impossible.
- 03
Preserves the dissent.
The arbiter rules, and the minority report is kept, not erased. Losing the vote does not mean leaving the record.
mechanism: three sheets written independently, isolation enforced
Status: every phase is built and tested. 322 tests, replay-complete traces, a deanonymization probe and a benchmark harness all exist. What does not exist is a single number produced by a real model: every eval runs against mocks, which proves the harnesses work and proves nothing about the models. The benchmark table goes in the README the day someone runs it with real keys, whatever it says.
Read the code →position-change target: 15-60%
0% = theater · 100% = herding
Governor
An agent can stay under budget on every step and still burn the night retrying one failing call. The limit that matters is not how long a run is. It is whether the run is going anywhere.
evals/runaway_eval.py: 11/11 · retry storm caught at step 2 · healthy 150-step run never left OK
A runaway-session governor.
Governor watches surface behavior, repeated text, repeated calls, repeated errors, growing prompts, and moves a five-rung ladder: OK, WARN, THROTTLE, PAUSE, KILL.
- 01
Dumber than what it watches.
Every detector is deterministic: token counts, call signatures, state digests. No model is called anywhere in the safety path.
- 02
Escalates fast, descends slow.
Evidence arrives all at once, so escalation may jump rungs. Coming back down is earned one rung at a time, after consecutive healthy steps.
- 03
Evidence, not authority.
A model asking to be killed reaches a human, never the kill switch. Kills destroy state, and most loops are recoverable if caught early.
Escalation is immediate and may jump rungs. Evidence arrives all at once.
{"step":1,"event":"tool_call","id":"‹masked›"}These are protocol illustrations, not live sessions. Every number in them matches the shipped eval scorecard (11/11 scenarios passing in CI). Code: github.com/JoaquinDG/governor
The receipts, executable
Everything claimed on this page runs offline on your machine. No API keys needed.
$ git clone https://github.com/JoaquinDG/switchboard && cd switchboard$ PYTHONPATH=src python3 -m unittest discover -s tests # 212 tests$ PYTHONPATH=src python3 evals/routing_eval.py # 8 routing scenarios$ PYTHONPATH=src python3 examples/quickstart.py # full demo, no API keys
How this was built
These systems were built the way they say AI should be run: specs before code, each repo started as a PRD with goals, non-goals, and rejected alternatives; evals that caught real bugs before launch, Switchboard's scenario suite caught its gates failing open and cost pressure routing work to underqualified models, documented in the README and now regression tests; Governor's audit found six defects where the safety property was stated in a docstring and enforced nowhere, all eight findings documented; and multiple AI agents dispatched in parallel and supervised with different models building different systems, coordinated like the trilogy itself prescribes.
The portfolio is the product demo. I didn't just build a dispatcher, a deliberation protocol, and a watchdog. I ran my own build process on them.