How Brain shapes every agent your team uses
Brain is more than a chat surface. The real leverage is what happens automatically: every agent your team runs gets the org's relevant Brain context injected into the system prompt, so outputs are brand-aware without any per-rep configuration.
Which agents pull from Brain
| Agent | Brain categories used |
|---|---|
| Outreach drafts (Origin) | voice_tone + sales_sop |
| Account briefs (Account Prospecting) | sales_sop + procedure |
| Customer Success suggestions | All four (sales_sop, voice_tone, procedure, deliverable) |
| TOV Database scoring | voice_tone |
| ICP scoring | sales_sop |
| Direct Brain chat at /brain | (every category, retrieved freshly per query) |
What "auto-injection" actually means
When an Enterprise org member runs an agent, the agent's wrapper:
- Identifies the user's organization.
- Pulls top-K relevant chunks from the org's Brain — biased toward the categories that matter for this task type.
- Prepends them to the system prompt as a
## ORG CONTEXT (cached)block, marked with cache-control so the LLM doesn't re-process the same prefix on every successive call within a session. - Runs the model.
- Records which Brain docs were referenced into
brain_document_citationsfor the coverage analytics.
The user sees no UI difference. They just see the output start sounding like your brand instead of a generic LLM.
Caching + cost
Brain context is heavily cached. The cache key is roughly (org_id, task_type, retrieved_chunk_set). As your team works through a session, the prefix gets cached after the first hit and subsequent calls pay the cache-read rate (a fraction of input rate).
Brain context auto-injection does not add a per-call charge. You pay for the underlying agent task (6 units for outreach, 15 for an account brief, etc.) — the Brain context is bundled in.
What doesn't get auto-injection
- Background workers (catalyst correlation, auto-ICP scoring) — these run at scale; we skip Brain to control cost. Manual invocations of the same agents do pull Brain.
- Enrichment (email/phone/LinkedIn lookup) — these aren't generative; no system prompt to extend.
Failure modes
If Brain retrieval fails for any reason, the agent runs without Brain context. The user-facing call always succeeds; the only signal is a WARN log line. We made this an explicit guarantee so a hiccup in retrieval never breaks an outbound draft mid-workflow.
Verifying it's working
Two ways to verify Brain context is influencing your team's outputs:
- A/B compare. Have a Pro user and an Enterprise org member generate the same outreach draft for the same prospect. The Enterprise output should reference your brand voice / SOP language; the Pro output should be generic.
- Coverage analytics. Visit the admin Coverage tab on
/brain(Stage 5). Each Brain document shows a citation count from the last 30 days. If the SOP and voice/tone docs are getting cited heavily and the others are at zero, that's expected. If one of your "important" docs has zero citations, it's either filed in the wrong category or the content isn't matching the kinds of questions your team's agents are answering.
Ask it on the help page and the assistant answers from the whole corpus. If it cannot, team@dealarena.io can, and the answer usually becomes a doc.
