Company Brain boundary

Type: agentic-system-analyses/types/agentic-system-boundary.md

Boundary and evidence

This analysis covers the self-hostable Company Brain application, not merely its memory search API. Its intended use is to explain how the shipped Slack agent admits requests and observations, retains company knowledge, later retrieves it, acts through connected tools, and maintains its state. The frozen revision is ef8a45ef3aed97b32599c7d8007edefab6d14e9a; the applicability cutoff is 2026-10-14. The overall evidence tier is code-grounded. HEAD was checked against the acquired commit, and the checkout was inspected read-only.

Classification: enclosing runtime; whole-system. Company Brain owns the application-level model/tool loop and its Slack and scheduled callers. Cloudflare and the Agents SDK supply infrastructure underneath that runtime; their presence does not make the shipped Slack driver an excluded enclosing application. The deployed worker exports the Durable Object, mounts the shipped HTTP routes and serves the shipped UI (SRC-1, src/worker.ts, wrangler.jsonc). Slack events dispatch into the organization's agent; the model loop selects the model, messages, tools and stopping conditions (SRC-1, src/routes/slack/index.ts, src/brain/turn/agent.ts, src/brain/turn/loop.ts). Minimum implementation anchors for this classification:

export class CompanyBrainAgent extends Agent { --- src/brain/turn/agent.ts

return args.deps.streamText({ model: args.deps.getModel(args.profile.name, args.env), system: args.system(), messages: initialMessages, tools: args.tools, activeTools: resolveActiveTools(), abortSignal: args.abortSignal, stopWhen: [ args.deps.stepCountIs(limit), () => args.suspendRequested?.() === true, ], --- src/brain/turn/loop.ts

The functional boundary includes installation and authentication, event admission, explicit and passive Slack turns, turn coordination and finalization, approvals and access leases, tool discovery and execution, sandbox integration, memory capture and retrieval, organization and personal configuration, skills, signup research, operator-triggered research, scheduled tasks and administration. Shipped callers remain inside the boundary even when they supply state to computeTurn rather than implement the loop themselves (SRC-1, src/routes/index.ts, src/brain/slack/turn.ts, src/brain/tools/scheduling.ts, src/brain/auto-research/draft.ts).

The memory loop was traced across its shipped producers and consumers. Completed interactive Slack turns call writeMemories; channel observation uses a model to select durable clusters and persists a batch before writing it. writeMemory passes documents through the compatibility adapter, registers tags and tree mappings, and initiates node splitting. Read scope is resolved before semantic search, and search_company_brain calls that search from the turn's tool set. Ambient profile, interaction style, workspace prompt and skills are loaded into turn context. Delayed post-turn observation writes team-level operating/style notes to the agent-self container, while research planning and scheduled turns consume retained context. These paths are included rather than treated as external maintenance (SRC-1, src/brain/slack/turn.ts, src/brain/slack/channel-observe.ts, src/brain/memory/index.ts, src/brain/memory/read-scope.ts, src/brain/memory/search-brain.ts, src/brain/turn/tools.ts, src/brain/turn/compute.ts, src/brain/turn/post-turn-reflect.ts, src/brain/turn/interaction-observe.ts, src/brain/memory/split.ts, src/brain/auto-research/plan.ts, src/brain/tools/scheduling.ts). This establishes scope and initial loop coverage, not the correctness of every permission or stale-write safeguard.

External dependencies and evidence limits. Supermemory's hosted service, model providers, Slack, remote MCP servers and connected applications, Cloudflare infrastructure, and optional Daytona infrastructure are outside the implementation source boundary. Their local clients, adapters, policies, deployment configuration and container recipe remain included. The memory adapter explicitly delegates processing to the external service:

• Write one document into supermemory. Extraction, chunking and embedding all
• happen server-side, so this is a single API call wrapped as an Effect so the
• original call sites and their error handling stay intact. --- src/compat/routes/memories/handler-effect.ts

The inspected adapter calls client.documents.add; the memory client constructs a Supermemory SDK client from the deployment key (SRC-1, src/compat/routes/memories/handler-effect.ts, src/memory/client.ts). No service implementation or frozen remote response is supplied. Consequently this run can establish the application's requests and treatment of returned data, but cannot establish the service's extraction, deduplication, supersession, embedding, ranking or retention behavior internally. Similarly, source-level tool and approval policy cannot establish provider compliance, successful remote actions or end-to-end privacy in a deployed workspace. Dependencies listed in package.json and bun.lock are not separately acquired sources; only the shipped local patches are part of this implementation boundary (SRC-1, package.json, wrangler.jsonc, sandbox/Dockerfile).

No observed run or causal experiment was supplied. Tests, delivery guards, research draft flags and observability code are included as authored implementation/evaluation surfaces, not evidence that a test passed or a production outcome occurred (SRC-1, package.json, test/workers-shim.ts, src/brain/slack/triage.test.ts, src/brain/codemode/quickjs-executor.test.ts, src/db/migrations.test.ts, src/brain/observability/reply-quality.ts, src/brain/auto-research/draft.ts). The target, tests and examples were not executed.

Shipped documentation is eligible as doctrine/design, including declared operation, but is not authoritative over the current implementation. Some architecture references still name a prior monorepo and Postgres deployment; this run instead pins the current worker and D1 configuration (SRC-1, docs/architecture.md, src/worker.ts, wrangler.jsonc). The documentation itself warns:

spec.md predates recent work and still has stale passages, so treat neither as authoritative over the code. --- docs/README.md

Top-level tracked coverage

Every top-level tracked entry at the frozen commit is classified below. “Included” means eligible within the functional boundary, not exhaustively read. “Not initially inspected” denotes available material, not an access gap. Later analysis may inspect it at the same pin. Directory rows identify material paths, including the shipped state producers, maintenance paths, consumers and evaluation surfaces.

The whole-system designation covers the shipped application responsibility loop; it does not claim exhaustive line-level inspection or include remote service internals. No material shipped path was found inaccessible. Available but unread leaf files above are follow-up evidence, not grounds for describing their mechanisms now.

Source register