How Symbion works

1 · Capture

A small daemon runs on your machine and records each session as it happens — prompts, per-turn timing, tool calls, file touches, subagents and cost. It reads only what's newly appended, so it stays cheap, and it can never break the session it's reading.

The same model covers Claude Code, Codex, Cursor, Copilot and Antigravity, so a task in one tool is comparable with a task in another.

2 · Segment into tasks

A session is rarely one job. Symbion splits it into the distinct tasks actually pursued — interleaving allowed — and gives each an AI-written title and summary. Tasks, not sessions, are the unit that gets scored, because that's the unit you'd recognise as a piece of work.

3 · Score the direction

Each task is scored 0–100 on how well the human directed the agent:

• Goal clarity — was the objective, the constraints and the definition of done stated up front?
• Decomposition — was the work broken into steps the agent could act on?
• Redundancy & correction — how much rework, restating and course-correcting was needed?
• Token & time efficiency — was the outcome reached without waste?

Every judgement cites the turns it came from. When the signal is too thin to be fair, the judge abstains rather than inventing a number — an honest gap beats a confident guess.

4 · Coach the prompt

The prompt is the only lever you actually control, so it's the coaching atom. Symbion shows patterns across your prompts, and on any item: your prompt → a better prompt → why. Individual prompts are never given their own score — patterns are useful, per-prompt grades are just noise.

5 · Codify and propagate

A lesson that stays in a dashboard changes nothing. Symbion turns recurring findings into managed rules, skills and memory that your agents actually read — and, on team tiers, pushes them from the org to every member. That's the Loop closing.

How it stays fair

• It scores how you direct the work, not the hours you put in.
• It tracks your own trajectory over time, not a ranking against colleagues.
• When the signal is thin, it says so — estimated figures are labelled, and the judge abstains rather than guess.