Open Executive — Architecture
A reference map of how Open Executive is built — the components, data flow, and invariants of the running system, from the Executive orchestrator and its specialist council to the knowledge, memory, scheduling, and integration layers.
Diagram legend
Without the Jargon: What It Is
The 90-second version. A virtual executive with a council of specialists.
Without the Jargon: How It Decides To Act
Three modes — act on its own, propose for approval, or escalate to you.
Without the Jargon: When You're Not Watching
Morning brief, check-ins, nudges — Open Executive doesn't wait to be asked.
Without the Jargon: Who Approves What
Departments, heads, and approval tags — how proposals find the right person.
System Overview
High-level topology: clients, API, orchestrator, specialist agents, knowledge layer.
Request Lifecycle
Full round-trip of a single chat message, including tool-use loop and parallel specialist calls.
Agent Council
The Executive orchestrator and its specialist sub-agents — roles, models, and domains.
Prompt Caching
How the system prompt is partitioned for Anthropic prompt caching, and what breaks the cache.
Knowledge & RAG
ChromaDB collections, per-specialist domain filtering, and where retrieved context is injected.
SME Knowledge Review
The pending-review queue, priority ordering, and how rejected/approved items affect retrieval.
Memory System
Episodic SQLite memory (decisions, initiatives, advice, scheduled actions) and how it’s surfaced.
Peer Memory (Person + Department)
External peer-keyed memory. Per-person scope keyed by Person.id for cross-channel continuity, and per-department scope keyed by department_<slug> for institutional voice. Dialectic prefetch, fire-and-forget sync, peer-graph cross-pollination, per-fixture workspace isolation.
Org Structure
Departments, goals, checklists, cadences; people registry; authority gates and channel resolution (Discord/Telegram/email).
Audit Log
Searchable, append-only record of chat turns, specialist consults, tool calls, scheduled actions, alerts, and inbound integrations.
Data Schemas
Key Pydantic models and database tables — the shape of the data flowing through the system.
Workflows
Deterministic, multi-step orchestrations that produce structured artifacts, including the wait-for-human pause primitive.
Scheduler & Cadences
The async scheduler runner, cadence DSL, scheduled_actions, and proactive nudges through the user’s last-used channel.
Integrations
External channels (Slack, Discord, email, Telegram, Google Chat, MCP gateway), attachments, and how they connect.
External Monitoring
Polls the watchlist (vendor status pages, RSS / Atom feeds, stock tickers) and routes qualifying signals through the same alert pipeline as inbound email / Slack. Watchlist editable from chat.
Today / Morning Brief
The /today route — per-department goal health, a roster with awaiting-action counts, and proposals routed to a person.
API Reference
The FastAPI HTTP surface — endpoints grouped by router.
MCP Server
Open Executive exposed as an MCP server — company context as resources and the specialist council as tools, over Streamable-HTTP at /mcp for external agents.
User Guide Surface
The /guide page — plain-language, per-feature overviews served from static prebuilt JSON, sharing this page's loader and renderer but separate from this technical reference.
Talent / Executive Search
The in-house hiring vertical: engagements (searches), candidates, the ChromaDB matching graph, and the draft-and-approve recruiting workflows — surfaced in the /talent UI.
Staff Onboarding
Role-tailored onboarding for new hires: reusable templates, per-hire plans with phased task checklists, the role_onboarding brief workflow, a bounded ramp drip, and chat + /today integration.
Client Companies (Slots)
Multi-client mode for fractional executives: named save files of the full company context, one active at a time, with save-back switching and per-client MCP tool configs.