AGT-901 — Pipeline Brain
Layer 9: Reasoning · Tier 2 Brain Agent · LLM-native · Reads Tier 1 services · Never writes canonical data · Operator-invoked
L9 · Agent 01
Specced · v27
Tier 2 · Brain Agent
RevOps-owned
First true Agent in the OS. AGT-901 is not a GTM Service — it is a Brain Agent. The distinction matters: Services are deterministic, scheduled, and own canonical tables; Brain Agents are LLM-native, operator-invoked, and never write canonical data. Tier 1 services run fine if AGT-901 is offline. AGT-901 cannot run if Tier 1 services are stale or unbuilt — which is why L8 was promoted to Built before this layer was specced.
Purpose
AGT-901 is the cross-functional reasoning layer for pipeline-level questions: why a segment is off plan, what play would close the gap, which existing levers apply, where the system is missing coverage. It reads from Tier 1 service tables, optionally calls Tier 3 specialist tools, and produces analysis with full source-trace metadata for human review. It does not run on cadence; it is invoked by RevOps or sales leaders posing open-ended queries, and by the WBR/MBR narrative jobs (see AGT-704 charter update in v26).
Scope — what AGT-901 answers, what it doesn't
| In scope | Out of scope |
|---|
| "Why is Q3 commercial pipeline soft?" | Computing pipeline metrics (AGT-702 owns) |
| "What play would close the coverage gap in Mid-Market enterprise vertical?" | Setting quotas (AGT-101 owns) |
| "Which existing AGT-203 ABM motions could be redirected to fill the SDR pipeline gap in EMEA?" | Routing leads (AGT-202 owns) |
| "What's changed in the win/loss pattern this quarter that should inform next quarter's plays?" | Computing win/loss (AGT-703 owns) |
| Drafting a candidate sales play for human co-definition in the workspace | Activating a sales play (requires human promotion gate; AGT-302 executes) |
| Generating narrative sections of WBR/MBR/QBR (AGT-704 narrative call-out) | Generating metric sections (AGT-704 pulls fresh from MetricsCalc, never from brains) |
The hard line. AGT-901 never recomputes a number that a Tier 1 service owns. If MetricsCalc is stale, AGT-901 surfaces the staleness in its output and refuses to estimate. Same posture as AGT-704's "never recompute" principle — the brain is a synthesis layer, not an alternative source of truth.
Read contract — Tier 1 services consumed
AGT-901 reads from brain-ready views of Tier 1 service tables. Each view is a precomputed projection optimized for LLM consumption (10× input compression typical, no full table dumps). Brain-ready views are owned by the underlying Tier 1 service and are read-only from AGT-901's perspective.
| Tier 1 service | Brain-ready view | Used for |
|---|
| AGT-702 GTM Health Monitor | MetricsCalc.brain_view — Magic Number, NRR, GRR, CAC Payback by segment + trend | Plan vs actual diagnosis. "Are we off because of efficiency, ACV, or volume?" |
| AGT-401 Deal Health Monitor | Opportunities.brain_view — top-N at-risk deals by ACV × risk score | Identifying which late-stage deals are dragging coverage |
| AGT-402 Forecast Adjuster | ForecastLog.brain_view — current forecast + delta vs prior week + components (new logo / renewal / expansion) | Forecast bias diagnosis |
| AGT-201 ICP Scorer | Accounts.brain_view — tier distribution + scoring patterns by segment + vertical | "Is the ICP profile shifting? Are we under-tiered in a growing segment?" |
| AGT-206 Account Prioritization | AccountPriorityScore.brain_view — composite distribution + queue freshness | "Are reps working the right accounts?" |
| AGT-105 Sales Capacity Planning | CapacityPlan.brain_view — deployed eRep + ramp status + open territories | Coverage gap diagnosis — structural vs situational |
| AGT-703 Win-Loss & Forecast Accuracy | WinLossLog.brain_view — trailing 4-quarter loss reasons + forecast bias by segment | Pattern-level diagnosis of execution issues |
| AGT-303 Cadence Intelligence | CadenceEventLog.brain_view — sequence performance + suppression patterns | "Is outbound quality the bottleneck, or is it lead supply?" |
| AGT-604 Voice of Customer | VoCSynthesisLog.brain_view — trailing 90-day themes + competitive mentions | "Are customers telling us something we're not hearing in win/loss?" |
If a Tier 1 service has not yet shipped a brain-ready view, AGT-901 falls back to reading the canonical table directly with explicit token-budget enforcement — but this is a temporary fallback, not the design target. Brain-ready views are the contract.
Write contract — what AGT-901 may write
| Target | What's written | Canonical? |
|---|
BrainAnalysisLog | Every query: question, sources read with timestamps, output narrative, proposed actions, confidence flags, model used, token cost | No — this is the brain's own log, not a canonical OS table |
SalesPlayLibrary (status: draft) | Proposed play definitions: name, hypothesis, target segment, target accounts, suggested cadence, success criteria, source-trace lineage | No — drafts only. Promotion to active requires human approval (see Promotion gate below) |
| WBR/MBR/QBR narrative sections | Per AGT-704 v26 charter: narrative content for sections 1 and 7 of MBR/QBR. Metric sections are off-limits. | No — written into BusinessReviewLog narrative fields, not metric fields. AGT-704 still owns the artifact and gates publication. |
| What AGT-901 may NEVER write |
|---|
Opportunities / ForecastLog / MetricsCalc | Anything — canonical pipeline/forecast/metric data is Tier 1 only | n/a |
QuotaStore / CompPlans / RevenueRecognitionLog | Anything — fiduciary data must remain deterministic | n/a |
ABMPlaybook directly | Anything — promotion from SalesPlayLibrary to ABMPlaybook is a human-gated action that AGT-203 executes after approval | n/a |
Promotion gate — from draft play to active play
A core architectural commitment: the brain proposes, humans co-define and approve, then AGT-302 executes. The volume cap is the safety mechanism — brains cannot quietly proliferate plays.
| State | Who can transition | What's required |
|---|
draft | AGT-901 (write) | Brain produces a play proposal; lands in workspace for human review |
under_review | RevOps or sales leader (manual) | Human picks a draft to refine into a candidate active play. May edit hypothesis, target, cadence in the workspace. |
active | SLM + RevOps (joint approval) | Both must approve before activation. AGT-302 reads active plays only. Hard volume cap: 3–8 active plays per segment per quarter (configurable per segment). |
retired | RevOps | Play removed from active execution. AGT-302 stops new sequence generation; in-flight sequences complete. Outcomes recorded for retrospective. |
Hard volume cap is enforced at promotion time, not at proposal time. AGT-901 may produce many drafts; only a small curated set ever reaches active. This matches the operating model in the v26 architecture eval: too many plays overwhelm reps and managers; sales leaders only commit to plays they helped define.
Use case patterns
| Pattern | Trigger | Sources read | Output |
|---|
| Plan diagnosis | RevOps query: "Why is [segment/period] off plan?" | MetricsCalc, ForecastLog, Opportunities, CapacityPlan brain-views | BrainAnalysisLog narrative with cited sources, decomposition (efficiency vs ACV vs volume), suggested levers |
| Coverage gap proposal | RevOps query: "Mid-market commercial vertical is under-covered. Options?" | Accounts, AccountPriorityScore, CapacityPlan, ICP scoring brain-views | BrainAnalysisLog with diagnosis + 1–3 candidate plays drafted in SalesPlayLibrary |
| Quarterly play refresh | Pre-planning kickoff (operator-triggered, not cadence) | Trailing-quarter outcomes from existing active plays + win/loss patterns + VoC themes | BrainAnalysisLog retrospective on prior plays + 2–5 draft plays for next quarter |
| WBR narrative | AGT-704 invokes for MBR section 1 (month narrative) | 4 weekly WBR artifacts + MetricsCalc month summary | Narrative content written to BusinessReviewLog narrative fields per AGT-704 charter |
| Anomaly explanation | RevOps query after AGT-702 breach alert: "What's driving the NRR drop?" | ChurnRiskLog, ExpansionLog, CustomerHealthLog, UsageMeteringLog brain-views | BrainAnalysisLog narrative + cohort decomposition with cited sources |
Output structure — every BrainAnalysisLog row
Every output AGT-901 produces is a BrainAnalysisLog row with full source-trace metadata. This is how the layer earns audit-grade reasoning out of a non-deterministic substrate. See BrainAnalysisLog production schema for field-level detail.
- Question — the operator query verbatim, or the AGT-704 narrative-job prompt.
- Sources read — structured list of (table_name, view_name, last_refresh_timestamp, row_count_consumed). Every numerical claim in the output ties back to one of these.
- Narrative output — the brain's prose response, with inline source citations matching the sources-read list (e.g.,
[src:1], [src:2]).
- Proposed actions — structured list of (action_type, target, lever, justification). Action types include
draft_play, flag_coverage_gap, recommend_query_for_human, none.
- Confidence flags — per claim:
high_confidence (single source supports), multi_source (multiple corroborating sources), inference (reasoning beyond direct source), speculation (the brain explicitly marks where it's guessing).
- Source freshness gate — if any source was stale at read time, the output is flagged
data_staleness_acknowledged and the staleness is surfaced in the narrative. Per AGT-704 staleness model.
- Cost metadata — (model, input_tokens, output_tokens, cached_tokens, cost_usd_estimate).
Cost guardrails
Per the v26 architecture eval cost model. The brain is operator-invoked, not on cadence — cost scales with usage, not with account count.
| Tactic | Implementation |
|---|
| Default model | Claude Sonnet tier. Opus reserved for explicit "deep analysis" mode where the operator opts in. |
| Prompt caching enabled | 5-minute TTL cache on system prompt + brain-ready views. Iterative queries on the same context pay the diff only. |
| Brain-ready views, not full tables | 10× input compression typical. Tier 1 services own the view definitions; AGT-901 is the consumer. |
| Per-query budget cap | Hard limit: 50K input + 5K output tokens per query. Queries exceeding the limit return a budget-exceeded result instead of partial output. |
| Monthly tier budget alert | Alert at 75% of monthly budget (RevOps-configurable, default $500/mo for Pipeline Brain). Hard stop at 100% requires manual override. |
Eval criteria
Per the v26 architecture eval directive: build the offline eval harness before turning the brain on in production. AGT-901 has its own eval suite separate from AGT-902.
| Criterion | Measurement | Pass threshold |
|---|
| Source citation rate | % of numerical claims with valid source citation pointing to a real Tier 1 row | ≥ 95% |
| Hallucination rate | % of outputs containing claims not supported by cited sources | ≤ 2% |
| Staleness recognition | When a source is stale, % of outputs that surface the staleness vs. silently using stale data | 100% — hard requirement |
| Diagnosis accuracy | 30 historical "why off plan" questions where retrospective ground truth exists; % where brain's top-2 causes match | ≥ 70% |
| Play survival rate | % of brain-drafted plays that survive co-definition + reach active state | ≥ 30% (playbook composition matters more than draft volume) |
| Active play outcome lift | Win rate / sales cycle / ACV on brain-co-designed plays vs. plays without brain involvement, cohort-level | Track quarterly; positive lift expected by Q+2 after launch |
Failure modes
| Symptom | Likely cause | Action |
|---|
| Brain fabricates a number not in cited sources | Hallucination — model failure | Caught by eval suite. Flag in BrainAnalysisLog. Pattern triggers prompt revision + model swap consideration. |
| Brain cites a stale source without flagging | Staleness check bypassed | Hard fail. Treated as a sev-2 incident: brain offline pending fix. Source-trace integrity is non-negotiable. |
| Many drafts, low promotion rate | Brain not aligned with sales leadership's actual play criteria | Co-definition working session: refine system prompt + add few-shot examples from approved plays. |
| High promotion rate, weak cohort outcomes | Plays sound good in workspace but don't perform in market | Quarterly retrospective. May indicate brain is over-fitting to leadership preference rather than underlying signal. Tighten eval suite. |
| Cost overruns | Caching disabled, full tables passed instead of brain-ready views, or excessive Opus use | Audit recent queries. Enforce brain-ready view requirement. Reset Opus opt-in. |
| Tier 1 staleness propagating | Brain reading a stale view | Brain refuses to produce output without staleness flag. RevOps triages the upstream service per AGT-704 staleness model. |
Invocation paths
| Path | Frequency | Initiated by |
|---|
| Operator query (chat-style interface) | Ad-hoc | RevOps, sales leader |
| WBR/MBR/QBR narrative job | On-cadence with AGT-704; AGT-704 calls AGT-901 for narrative sections only | AGT-704 |
| Quarterly play refresh batch | Once per quarter, kickoff of planning | RevOps (manual trigger) |
| Anomaly explanation | Triggered by RevOps after AGT-702 breach alert — not auto-fired | RevOps |
No cadence-driven invocation. The brain does not run on a schedule. This is intentional — operator-driven invocation keeps cost bounded and aligns with the "humans in the workspace, not autonomy" approval model.