Ganglia is a multi-agent thinking system. You submit a question. Specialised agents research it, structure it, attack it, and synthesise an answer, while a separate coach makes you reason through it yourself in parallel.
Most of the time, a single AI chat is the right tool. Ask a question, get a synthesised answer, move on. The trouble starts on questions where you already hold a view but are not sure it is right. A single model tends to agree with you, hedge, or blend every perspective into a safe average.
Ganglia is built for that narrow, high-stakes case. It splits the work across agents with sharply different jobs. One finds prior evidence. One structures the problem. One attacks the emerging frame. Named persona critics bring genuinely different priors. A meta-reviewer keeps them honest. A final agent takes a position rather than summarising the noise.
Running underneath the whole thing is a second purpose: to keep your own reasoning sharp as you lean on AI more. A separate coach runs your thinking in parallel and refuses to hand you the answer.
One question in, one synthesis out. Between them sit five phases, coordinated by an orchestrator called the Cortex. Key: agent artefact you
The Reader classifies your raw prompt: question type, what you have stated versus what you are assuming, and where the gaps are. The Proctor then interviews you to fill those gaps and produces the Brief, the sharp statement of the question every other agent works from. Nothing runs until you approve the Brief.
The Cortex decides which agents to convene for this specific question and tells you why, one line each, before anything runs. The Forager, Architect, and Weaver always run. The Contrarian, the Fellows, and the Editor join based on the Brief. The narrative agents run only on request.
On strategic and design questions, the Contrarian's critique goes back to the agents it targeted. Each gets one chance to concede, defend, or refine in 200 words or fewer. One round only. This is the bet that a single round of structured disagreement improves the output, and the system watches whether it adds signal or just noise.
The Editor reviews every output for role drift, vagueness, and safe-but-boring thinking, and sends weak ones back once. The Weaver then takes the approved set and writes the Synthesis: not a blend, a position, with what is settled and what is still open made explicit.
You read the Synthesis, plus any narrative outputs. The full session, every agent output included, is captured as a Trace in the Vault. You are offered the chance to promote the passages worth keeping into your long-term second brain. Then the session closes. One prompt, one session, no silent chaining.
The name is deliberate. A ganglion is a cluster of nerve cells; the ganglia are the agents; the Cortex coordinates them. Every role earns its place, and none of them is allowed to do another's job.
Coordinates the session. Decides which agents run, in what order, when to loop, when to stop, and when to hand back to you. Produces no content of its own.
Reads your raw prompt and produces the Read: question type, what is stated versus held versus assumed, and the gaps. It does not research or answer. It reads.
Interviews you to fill the gaps and produces the Brief. Asks specific, load-bearing questions in batches, then proposes the Brief for your approval.
Runs your own creative process in parallel using a structured technique (pre-mortem, assumption reversal, first principles, and others). It never offers its own answer, even when asked. A nudge unsticks you; a rescue would rob you of the thinking.
Finds prior context and external sources. Checks your own second brain first, project notes second, the web third. Extracts what the Brief needs, cites precisely, and flags where the material is thin rather than papering over it.
Structures the question into the distinctions, decomposition, and dimensions the session works within. It frames the problem; it does not answer it or propose a specific solution.
Attacks the emerging frame. Leads with the single most load-bearing weakness, names what would change its mind, and stays specific. A generic objection is worse than no objection.
One to three named-perspective agents, chosen per question. Each applies a real thinker's methods and priors to the Brief, including what that thinker would advise against. Their job is to break the convergence a single model tends toward.
Reviews every agent's output before it reaches the Weaver. Checks each one did its job, stayed specific, and pushed hard enough. Sends weak work back once. It edits; it does not rewrite.
Produces the Synthesis you read. Finds the best answer the agents have collectively reached, says who is right and why, and is honest about what remains uncertain. Includes a considered-and-rejected note so you know which paths not to revisit.
Casts the session's position as a short story, case study, or allegory, because stories stick where briefs do not. Works from the raw material, not the Synthesis, to keep its narrative independent.
Takes the position as wrong and tells the most plausible story of how it fails, then names the early warning signs. A position that survives being seriously imagined as wrong is stronger than one that has not.
Ganglia names its artefacts so a session can be discussed precisely.
Your raw input. The starting point.
The Reader's structured classification of what the system understands the question to be.
The load-bearing artefact. The sharp statement of the question, the held positions, the open questions, and the success criteria that every agent works from.
The final output you read. A position, not a summary.
The full record of one session: prompt, read, brief, every agent output, the synthesis, and any follow-up questions.
Where Traces are stored, so a session can be retrieved and built on later.
The act of moving a passage worth keeping from the Vault into your long-term second brain.
Roughly one question in twenty warrants the full machinery, possibly fewer. The friction is the point: it earns out only when the question genuinely benefits from the structure.
You do not always need the whole cast. Sparring Partner only when you want to think harder about something you already understand, such as pre-meeting prep. Architect and Contrarian when you have a frame you are committed to and want it attacked before you take it to others. The full system is reserved for the questions that genuinely need all of it.
Version 0.1 was tested on a genuine personal decision with several competing dimensions. The whole cast ran except the Sparring Partner. Here is what held up.
The fair test is not "Ganglia versus a single chat from a cold start". Anyone using this already iterates on hard questions in ordinary chats. The real test is whether a full session beats another round of the iteration you were already doing. On the first session, it did, and by a clear margin.
The red-team critic and the anti-narrative agent were rated the strongest contributors. The critic quantified the cost of being wrong and surfaced consequences the framing had hidden. The anti-narrative dramatised the failure path in a way pure analysis could not, which made its early-warning signals genuinely actionable.
Early breakages were about plumbing, not reasoning: the bootstrap was wrong about which data stores it could reach, and access patterns were inconsistent between config files. The persona agents also needed a hard-constraints check, after one produced an in-character plan that ignored the Brief's fixed dates.
The spine works: prompt to read to brief to dispatch to synthesis. The structured disagreement of the response pass added signal rather than noise. The honest counterfactual question got rewritten for future sessions. A successful test of version 0.1.
The agent prompts are first drafts being hardened on live sessions. The design is deliberately revised only after the first handful of real runs, not before.