Sessionsession.ad · community · academy · lexicon

A community, an academy, and a knowledge base for people who think for a living. For students. For engineers. For anyone who still believes the work matters more than the demo.

Edited by Secan98 · Last updated: , · Contents · Blog · Sign in · Create a free account

You are here: HomeAbout Session

Diagram: a question fans out to three agents, who debate, are scored by a small judge, and converge into a synthesised answer.
Nowessay 9 in the archive Currently writing"why the judge model should be small" EditionVol. I, No. IX Members1,247 free subscribers

Session is a project about how to think clearly with more than one mind in the room. It has three parts that work together: a free academy with structured lessons, a community for peer review and shared experiments, and a knowledge base & lexicon of terms and worked examples. The desktop tool sits at the centre, a canvas where you let several AI models debate a question and produce a single synthesised answer.

Reading the canvas, the disagreements, and the consensus is itself a skill. The Academy and the Lexicon exist to teach it. Everything below is open and free; an account simply lets you save progress, post in the community and keep a personal notebook in the portal.

1.Mission

Single-model chat hides disagreement. A model picks an answer, you read it, you trust or you don't. There is no record of what the second-best answer was, or who would have argued otherwise. Session treats that as a problem worth fixing. We bring several models to the same table, record what each one says, mark where they agree, and produce a final answer that is honest about what is contested.

This site is for the person who reads slowly. The student who looks things up. The engineer who wants to know why. The lifelong learner who never stopped being curious. If that is you, welcome home. There is no funnel here, no popup, no autoplay video, no cookie banner. Just essays, definitions, and a community of people who care about getting the answer right.

Around the practice of multi-agent reasoning we have built a small school. The Academy teaches the underlying ideas. The Community is where readers compare notes, share canvases, and review one another's runs. The Lexicon is the slow, accumulating vocabulary of the field, written carefully and corrected by members. Edited and written end to end by Secan98, fifteen years in IT, working from an old ThinkPad and a quiet office.

2.The Academy

Free, structured lessons. Each module is a short reading plus a worked example you can open inside the desktop tool. Designed so you can finish a track in a weekend.

Tracks (currently published)

  • Track A, Foundations of Multi-Agent Reasoning
    8 lessons · 2 h 40 min · reading-only · level: introductory

    Why one model isn't enough. What "agreement" means. Reading a debate transcript. The synthesis card.

  • Track B, Designing a Debate
    6 lessons · 2 h 05 min · level: intermediate

    Choosing agents, picking a synthesizer, writing prompts that elicit disagreement, judging convergence.

  • Track C, Local-First & BYOK
    4 lessons · 1 h 20 min · level: practical

    Running local models for offline debates, securing your keys, when cloud sync helps, when it doesn't.

  • Track D, Provider Comparison Studies
    5 case studies · reading-only · level: advanced

    Where frontier models disagree in practice, on code review, on long-horizon planning, on scientific summarisation.

3.The Community

Members can publish their canvases as read-only "notebooks" for others to read, comment on, fork, and re-run with their own keys. The forum is structured around three rooms.

Room: Studies Shared canvases, with the prompt, the providers, the transcript, and the synthesis. Comments allowed; one canvas per thread.
Room: Methodology Discussion of debate protocols, judge models, prompt-design patterns, replication notes.
Room: Help & Setup Practical questions: keys, providers, MCP servers, sync.

Posting and commenting requires a free account. Reading does not.

4.Knowledge Base & Lexicon

A growing dictionary of terms, written for newcomers but precise enough to cite. Each entry has a short definition, a longer note with examples, and cross-references. Members may suggest edits.

Agent
A configured language-model node on the canvas. One agent has one provider, one model, one role and one set of instructions. See also: Provider, Debate.
BYOK, Bring Your Own Key
The principle that API keys belong to the user and never leave the user's machine. Session stores keys in the operating-system keychain. See also: Local-first.
Debate
A round-robin protocol where agents respond to each other's outputs rather than producing one answer in isolation. Round 0 is parallel; rounds 1 and later pass each agent's structured turn to its peers. See also: Synthesis, Judge model.
Judge model
A small, cheap model whose only job is to score how much the agents have converged. Used to decide when to stop the debate early. See also: Convergence.
Local-first
A design where the canonical copy of your data lives on your machine. Cloud sync, if used, is an optional mirror, not the primary storage. Session is local-first by default.
MCP, Model Context Protocol
A standard way to give an agent real tools (filesystem, web fetch, repository, custom) so it can act, not just speak.
Provider
The company or runtime that hosts a model. Frontier providers and any OpenAI-compatible endpoint qualify; local runtimes count too.
Synthesis
The consensus answer produced by a designated synthesizer agent that reads the full debate transcript and writes a single, attributed response. Minority dissent is preserved inline. See also: Debate, Convergence.
Convergence (agreement %)
A heuristic measure of how close the agents have come to a shared answer, scored by the judge model. A debate may early-exit at ≥ 85%.

Full lexicon is available in the portal once you create a free account.

5.The Desktop Tool

The tool is what the Academy teaches you to use. It is a desktop application with a node-based canvas. You drop agents on the canvas, connect context to them, run a debate, and read the synthesis.

How a session unfolds

  1. Drop agents on the canvas. Pick a template, or wire your own. Each agent is one node, one provider, one model.
  2. Connect context. Brain edges from notes, documents, web links, repositories, MCP servers. Typed sockets refuse invalid wires.
  3. Run a debate. Round 0 fans out in parallel. Rounds 1 and later pass each agent's structured turn to its peers. Tokens stream; edges glow.
  4. Read the synthesis. The designated synthesizer reads the full transcript and emits a consensus card. The debate early-exits when agents converge.
A four-column diagram: round 0 fan-out, rounds 1+ round-robin critique, judge model with convergence threshold, synthesis card.
Fig. 1How a debate is structured inside Session. Black ink: control flow. Crimson: peer-to-peer critique. Dashed: early-exit loop.

Worked example

A canvas with three agents is asked whether to migrate a small project from one database to another. Two rounds of debate. Two agents quickly agree on scaling. The third introduces cross-device sync as an additional argument. The synthesizer writes a single answer with attribution; convergence is reported at 92%.

  prompt: should we adopt a different database?
  agents: agent-A, agent-B, agent-C
  rounds: 2  synthesizer: small judge
  ───────────────────────────────────────────
  synthesis: yes, on the agreed conditions.
    · A + B agree on scaling.
    · C adds: cross-device sync benefit.
    · minority dissent: none.
  convergence: 92%   tokens: 11,420   cost: ~cents

6.Recent Writing

The latest essays from the blog. Full archive at /blog.

  • Why one model isn't enough
    2026-05-04 · 9 min read · by Secan98

    A foundational essay on why a single confident answer hides the disagreement that hard questions deserve, and why putting several models in the same room is the smallest fix.

  • Anatomy of a debate protocol
    2026-04-29 · 10 min read · by Secan98

    How a multi-agent debate is actually structured: parallel fan-out, round-robin critique, a small judge, threshold exit, and a synthesizer that preserves minority dissent.

  • Local-first AI: a practical guide to Ollama-only debates
    2026-04-18 · 9 min read · by Secan98

    A hands-on guide to running three local models in a debate on your own machine. Hardware math, model selection, latency, and when to reach for the cloud.

  • BYOK and the case against centralized keys
    2026-04-12 · 8 min read · by Secan98

    Why an AI app that stores your model-provider keys server-side is a liability waiting to be discovered. The OS keychain is the only defensible default.

» All posts in the blog archive

7.Membership & Pricing

The Academy, the blog and the public lexicon are free to read without an account. A free account adds the portal: progress tracking, community posting, personal notebook, and the desktop tool with BYOK.

Plan Price What you get
Free, recommended for new members $0 forever Portal access, all academy tracks, public lexicon, community read+post, desktop tool with BYOK.
Pro $15 / month Multi-provider debate templates, MCP server attachments, cloud sync across devices, beta channel.
Team $25 / seat · min 3 Shared real-time rooms, SSO/SAML, audit log, admin dashboard, priority support.

Create your free account Sign in

8.Frequently Asked Questions

Do I need API keys for every provider?

No, only the providers you actually use. A single-agent run on the free plan needs just one. Multi-provider debate templates use whichever you have configured. BYOK means bring-your-own-key, keys live in your operating-system keychain and are never sent to us.

Where does my data live?

All sessions, canvases and run transcripts are stored locally by default. With Pro, you can opt-in to cross-device sync. Your API keys never leave your machine.

How does authentication work?

Two options. You can sign up with an e-mail address and a password (Argon2id-hashed, brute-force protected). Or you can use an e-mail one-time code. Either way, the desktop client receives a session token that signs you in for the web portal and the desktop app.

What is an "MCP server"?

Model Context Protocol, a standard way to give agents real tools (filesystem, web fetch, repository, custom). Pro lets you attach an MCP server to any agent so it can call tools mid-debate, with results visible to peers. See lexicon entry.

Can I self-host the backend?

Yes, on the roadmap for Q1 2027. Session is built around a desktop app plus an open-spec backend. Enterprise plans include a self-host licence.

Does this work without an internet connection?

It depends on the agents. Frontier providers all need the internet. Local runtimes work fully offline. The canvas, persistence and chat history are local-first; authentication and cloud sync need network.

9.Colophon & Contact

Session is a small project, hand-built with care. This document is plain HTML and one stylesheet; there are no trackers, no analytics, no third-party scripts. The font is Georgia. The author is Secan98.

Contact: hello@session.ad. Bug reports and corrections welcome. Pull requests for the lexicon live in the community room.


© 2026 Session · Home · Blog · About · Academy · Lexicon · FAQ · Contact
Hand-built with care · no cookies, no trackers, no third-party scripts.