How Kiln works

1 · The 60-second loop

sequenceDiagram
    autonumber
    participant V as Visitor
    participant O as Orchestrator (Opus 4.7)
    participant T as Tools (CRM · Vector · Exa)
    participant Subs as 5 Sub-agents (parallel)
    participant Slk as Slack #deal-desk
    participant Out as Artifacts (DOCX · PDF · XLSX · EML)

    V->>O: Run review on deal
    O->>T: Step 1 — fetch deal
    T-->>O: Deal record
    par Step 2 — gather context (parallel)
        O->>T: Vector k-NN over 40 deals
        T-->>O: 3 similar past deals
    and
        O->>T: Exa public signals
        T-->>O: 3-5 customer signals
    end
    par Step 3 — Promise.all dispatch
        O->>Subs: Pricing
        O->>Subs: ASC 606
        O->>Subs: Redline
    end
    Subs-->>O: 3 typed outputs
    O->>Subs: Step 4 — Approval routing (Haiku 4.5)
    Subs-->>O: ApprovalOutput
    O->>Subs: Step 5 — Comms drafts (Sonnet 4.6)
    Subs-->>O: CommsOutput
    O->>Slk: Block Kit post
    O->>Out: Generate 5 artifacts
    O-->>V: Synthesis (4-sentence verdict)

2 · The agents

Five specialist sub-agents under one orchestrator. Each lives in its own file under lib/agents/ with a markdown system prompt at lib/prompts/.

Why these models. The orchestrator and ASC 606 use Opus 4.7 — the synthesis verdict and the rev-rec reasoning are the two highest-stakes outputs and both reward the strongest model. Pricing, Redline, and Comms run on Sonnet 4.6 for a balanced cost/quality trade on math, clause analysis, and tone-aware drafting. Approval is a deterministic table lookup over the matrix, so Haiku 4.5 handles it at a fraction of the cost.

3 · Data flow

Three data sources feed the orchestrator before the first sub-agent dispatches: the committed SQLite CRM, the sqlite-vec virtual table of deal embeddings, and Exa for public customer signals. A fourth path — the cached-output store — replays prerecorded agent results for hero scenarios.

graph LR
    CRM[("SQLite CRM<br/>40 deals · 30+ customers")]
    Vec[("sqlite-vec<br/>1536-dim embeddings")]
    Exa[("Exa API<br/>public signals")]
    Cache[("cached_outputs/<br/>5 hero scenarios")]
    Orc{{"Orchestrator<br/>Opus 4.7"}}
    Pri["Pricing<br/>Sonnet 4.6"]
    Asc["ASC 606<br/>Opus 4.7"]
    Red["Redline<br/>Sonnet 4.6"]
    Apv["Approval<br/>Haiku 4.5"]
    Com["Comms<br/>Sonnet 4.6"]
    Slk["Slack<br/>kiln-demo"]
    Doc["DOCX · PDF · XLSX · EML"]
    Aud[("audit_log")]

    CRM --> Orc
    Vec --> Orc
    Exa --> Orc
    Cache -.cached replay.-> Orc
    Orc --> Pri
    Orc --> Asc
    Orc --> Red
    Pri --> Apv
    Asc --> Apv
    Red --> Apv
    Apv --> Com
    Com --> Slk
    Com --> Doc
    Orc --> Aud

    classDef store fill:#f5f5f5,stroke:#737373,stroke-width:1px,color:#0a0a0a
    classDef agent fill:#fafaf9,stroke:#0a0a0a,stroke-width:2px,color:#0a0a0a
    classDef out fill:#fff7ed,stroke:#c2410c,stroke-width:1px,color:#0a0a0a
    class CRM,Vec,Exa,Cache,Aud store
    class Orc,Pri,Asc,Red,Apv,Com agent
    class Slk,Doc out

4 · MCP server architecture

Tools aren't inlined into the agent code. They're exposed through four internal MCP servers — running in-process inside the Next.js app — that the orchestrator and sub-agents call by name. Tool names are namespaced mcp__<server>__<tool>.

MCP gives agents a clean tool boundary independent of where they run. The same agent code works inside Kiln's Node runtime, inside Claude Desktop, or against a future Clay-MCP integration — only the server registration changes. It also means a fork can swap the mock CRM for a real Salesforce backend by replacing one server file.

5 · Worked example — Anthropic Q1 expansion

What actually happens when the orchestrator runs on the Anthropic hero scenario — a $1.5M TCV strategic expansion with 5 non-standard clauses including an MFN. Timestamps are wall-clock from the live run captured during build verification.

1. T+0.0sOrchestrator

   Load deal from CRM

   mcp__crm__get_deal('deal_anthropic_2026q1_expansion') → full deal record. $1.5M TCV, 36-month term, 12% headline discount, 5 non-standard clauses (MFN, exclusivity, custom data residency, rollover credits, termination-for-convenience).

2. T+0.1sOrchestrator

   Step 2 — gather context (parallel)

   Fan-out to vector k-NN and Exa via Promise.all. Vector returns 3 similar deals (a 2025 Anthropic expansion, an OpenAI expansion, a Mistral mid-market deal). Exa returns 4 signals — recent funding, leadership change, new product launch, hiring signal. Step joins at ~3.2s.

3. T+3.3sPricing · ASC 606 · Redline

   Step 3 — dispatch sub-agents (parallel)

   Pricing identifies the headline 12% discount as 28.4% effective once the ramp and credits are factored in; proposes three alternative structures. ASC 606 identifies 4 performance obligations and flags the ramp as variable consideration requiring the expected-value method. Redline flags 5 clauses with concrete counter + fallback positions for each. Step joins at ~15s (ASC 606 is the long pole).

4. T+15.3sApproval

   Step 4 — approval routing

   Matrix evaluation triggers three rules: ACV > $500K → CFO; non-standard clause count > 3 → Legal; MFN clause present → CEO. Final chain: AE Manager → RevOps → CFO + Legal (parallel) → CEO. Estimated cycle time: 5 business days.

5. T+18.4sComms

   Step 5 — communications drafts

   Block Kit Slack post, AE briefing email (collaborative tone, walks the AE through the 5 clause counters), customer-facing email (collaborative, leads with alternative structures), and approval one-pager. All drafts cite the upstream outputs with concrete numbers, not paraphrases.

6. T+30.5sSlack + Artifact generators

   Slack post + 5 downloadable artifacts

   Block Kit message lands in #deal-desk in the kiln-demo workspace. Document generators emit redlined-msa.docx, order-form.pdf, approval-one-pager.pdf, ae-email.eml, and a 10-tab deal-summary.xlsx with live cross-tab formulas — open it in Excel and edit the discount cell to watch the rev-rec schedule recalculate.

7. T+~62sOrchestrator

   Synthesis verdict

   Four-sentence executive summary citing each sub-agent. Audit log row written. Review persisted to deal_reviews. UI receives the final SSE event and renders the verdict card + agent output cards.

6 · Caching strategy

7 · Output schemas

Every agent emits JSON conforming to a Zod schema declared in lib/agents/schemas.ts. Structured outputs are the reason the UI is renderable, the audit log is queryable, and the eval harness can grade runs at all.

// lib/agents/schemas.ts
export const PricingOutputSchema = z.object({
  list_price: z.number(),
  proposed_price: z.number(),
  effective_discount_pct: z.number(),     // real discount after ramp + free months
  margin_pct_estimate: z.number(),         // assumes 40% gross margin at list
  guardrail_evaluations: z.array(z.object({
    rule_name: z.string(),
    passed: z.boolean(),
    severity: z.enum(["info", "warn",
                      "block_without_approval", "block_absolute"]),
    actual_value: z.number(),
    threshold_value: z.number(),
    explanation: z.string(),
  })),
  alternative_structures: z.array(z.object({
    label: z.string(),                     // "Trade discount for term length"
    description: z.string(),
    proposed_price: z.number(),
    effective_discount_pct: z.number(),
    expected_acv_impact: z.number(),
    margin_pct_estimate: z.number(),
    rationale: z.string(),
  })).min(2).max(3),
  ltv_estimate_under_usage_assumptions: z.number().nullable(),
  similar_deal_references: z.array(z.string()),
  confidence: z.enum(["low", "medium", "high"]),
  reasoning_summary: z.string(),
});

{
  "list_price": 1800000,
  "proposed_price": 1530000,
  "effective_discount_pct": 28.4,
  "margin_pct_estimate": 22.1,
  "guardrail_evaluations": [
    { "rule_name": "max_discount_enterprise",
      "passed": false, "severity": "block_without_approval",
      "actual_value": 28.4, "threshold_value": 20,
      "explanation": "Effective discount exceeds 20% threshold once ramp + credits factored in." }
  ],
  "alternative_structures": [
    { "label": "Trade discount for 48-month term",
      "proposed_price": 1620000, "effective_discount_pct": 18.0,
      "rationale": "12 extra months reduce CAC payback to 14 months." },
    /* …2 more */
  ],
  "confidence": "high",
  "reasoning_summary": "Headline 15% becomes 28.4% effective…"
}

8 · Tech surfaces

The stack at a glance, for the technical reader who skimmed past the diagrams.

• Streaming·Server-Sent Events from app/api/run-review/[dealId]/route.ts; one event per agent step.
• Vector search·sqlite-vec extension on the same SQLite file. k-NN over 1536-dim embeddings.
• Embeddings·OpenAI text-embedding-3-small. Generated once at seed time; persisted as BLOB on the deal_embeddings virtual table.
• Document generation·docx for redlined MSA · pdfkit for order form + one-pager · exceljs for the 10-tab live-formula workbook.
• Slack·@slack/web-api with Block Kit JSON. Bot posts to #deal-desk in a real kiln-demo workspace; invite link surfaces in the sidebar.
• Agent framework·@anthropic-ai/claude-agent-sdk via query() per agent. Bounded reasoning (effort: low, maxTurns: 2) for predictable latency.