Brain and synthesis

Venture Brain converts operational observations into versioned, agent-specific working knowledge. Observations are the event stream; synthesis is the durable process that decides what should become a new Brain version.

KNOWLEDGE LIFECYCLE

Evidence becomes versioned working knowledge

1. 01

   Observe

   Workflows and people record useful evidence.

2. 02

   Filter

   Completion and importance gates protect quality.

3. 03

   Synthesize

   Eligible evidence is combined with the current Brain.

4. 04

   Version

   An immutable version becomes active for the agent.

Observations

An observation records something an agent or human action learned: a fact, decision, outcome, blocker, correction, escalation response, schedule change, or workflow event. Observations retain their venture and workflow context, importance, creation time, and synthesis state.

Only observations associated with completed known workflows are eligible; records from known incomplete workflows are withheld so partial or failed output does not become durable knowledge. Human observations are shared context and can trigger synthesis for existing agent brains even when no individual agent reaches its own threshold.

Observation storage is not the same as Brain memory:

• Observation: append-like evidence awaiting or supporting synthesis.
• Brain version: a synthesized snapshot for one agent.
• Active Brain: the version currently mounted/retrieved for that agent.
• Canonical source: strategy, repository, provider, resource, or project-management record that remains authoritative.

Eligibility thresholds

Each agent has an observation threshold that balances freshness against model cost. Strategic agents generally synthesize after fewer new observations; high-volume agents wait for a larger batch. Venture overrides can tune this behavior.

A lower threshold produces fresher Brain versions and more synthesis usage. A higher threshold reduces synthesis frequency but leaves new knowledge waiting longer. Use the value shown in Brain settings as the source of truth for the agent.

Only sufficiently important, eligible observations are included. Observations from incomplete or failed work remain out of the active Brain so unfinished output does not silently become durable guidance.

Schedule and manual sync

Each configured venture has a Brain synthesis schedule. The UI offers cadence choices from frequent synchronization through weekly and a Frozen position.

Freezing:

• pauses the venture schedule;
• excludes the venture from the global fallback synthesis run;
• disables manual “Sync all now”;
• stops freshness badges from aging beyond the freeze instant.

Freezing does not delete observations or Brain versions. Unfreezing resumes synthesis from accumulated eligible observations.

Model resolution

Brain synthesis uses one uniform provider/model pair for all agents in a venture run. The pair resolves from most specific to least specific:

1. venture Brain-synthesis provider and model;
2. organization Brain-synthesis provider and model;
3. platform Brain-synthesis default.

Both provider and model must be present at a tier. The resolver never combines a provider from one tier with a model from another.

The Brain schedule's model is separate from the general workflow LLM catalog. Changing a Developer or Content phase model does not change Brain synthesis. Use the Synthesis Model control in Brain settings.

Synthesis execution

One run resolves eligible venture/agent pairs, applies the selected model, and synthesizes each eligible agent within the venture context.

For each agent the synthesizer receives:

• the existing Brain snapshot, if present;
• eligible new agent observations;
• eligible human observations shared across agents;
• agent Brain configuration and importance thresholds;
• venture/repository context available to the synthesis workflow.

The result creates an immutable Brain version, makes it active for that agent, links the contributing observations, and records token/cost data. Shared human observations are considered complete only after the applicable agent syntheses finish.

A failed synthesis should not mark the observation set complete. Inspect the synthesis workflow and last agent result before triggering repeated manual runs.

Token usage and cost

Each agent synthesis reports input, output, cache creation, cache read, model, and estimated cost where the provider supplies usage. The business run aggregates those calls with the same token-pricing engine used by other workflows.

Brain synthesis cost is stored on each version and contributes to:

• per-agent Brain analytics;
• venture model-cost totals;
• month-to-date AI spend and the venture spend-cap comparison;
• organization token metering where configured.

Any execution compute is reported separately from model-token cost.

See Compute, tokens, and model selection for formulas and billing boundaries.

Brain analytics

The Brain surface can show:

• per-agent lifetime, seven-day, and 30-day observation counts;
• synthesis count and last-synthesized time;
• cumulative synthesis cost;
• median observation importance;
• a 30-day observation/synthesis time series;
• recent high-importance observations;
• Brain-version versus task-success correlation when enough history exists.

Correlation is evidence, not proof of causation. It becomes available only after the venture is old enough and Brain versions have correlated workflow outcomes.

Corrections and hygiene

• Correct the canonical strategy, resource, repository, or external system first.
• Record a human observation when the correction should influence agent behavior.
• Trigger manual synthesis only when waiting for the cadence is operationally costly.
• Raise thresholds for noisy, repetitive agents; lower them for strategic agents that need fresh decisions.
• Freeze when intentionally preserving a stable Brain during a sensitive operation or investigation.
• Do not treat Brain as a secret store. Credentials belong in encrypted resources and are provisioned at runtime.

What to expect

• Brain is working knowledge, not the canonical source for strategy, credentials, repository state, or provider records.
• A successful synthesis does not prove that every conclusion is correct; review important decisions and correct the canonical source first.
• Model, provider, or execution failures can delay a scheduled synthesis.
• Freezing preserves the current active versions while observations continue to accumulate for later processing.