Company OSArchitecture
Deep dive

Company OS,
layer by layer.

The simple version is four steps: connect, know, act, work. This page opens each one up, so you can see every part and how they fit together.

Click any part to jump to its details.

Part 01

Connect

How your data, rules and actions get in.

01Connect

Context engineering

Your data, your rules and your actions, plugged in.

Company OS connects to the systems you already run. It doesn't just copy data out. It learns your business rules too, and gets permission to act back in those systems.

  • Data: live sync from email, Slack, CRM, calendars, files and databases (800+ connectors)
  • Logic: your formulas, pricing rules, approval chains and existing scripts
  • Actions: safe write-back, like updating a deal or opening a ticket
  • Real-time, scheduled or streaming, with or without code
Back to the map
EmailSlackCRMDatabases
DataLive sync · 800+ connectors · Real-time
Pricing rulesApprovalsScripts
LogicRules · Formulas · Your models
Update a dealOpen a ticketPost a message
ActionsWrite-back · Webhooks · Streaming
Company OS
Parts 02–04

Know

How it understands your company.

02Know

Knowledge graph

One connected model of your company.

Every person, customer, deal, project and document becomes an object, linked to the others. The graph records what things are, what can be done to them, and the decisions made along the way.

  • Things: people, customers, deals, projects, documents
  • Actions: what each role or agent is allowed to do to each thing
  • Decisions: who decided what, when, and why
  • History: see any record as it was on any date
Back to the map
championrenewingworks onsigned inapplies toapprovedCUSTOMERAcme CorpPERSONJordan LeeDEALRenewal 2026PERSONMaria, SalesDECISION10% off year 2DOCUMENTContract v4ThingsDecisionsLinks, each with a meaning

Swipe to see the whole graph.

03Know

Documents & search

Every file, searchable by meaning, with sources.

Documents, images, audio and video are read, indexed and linked to the graph. Search mixes keywords, meaning and relationships, then checks its own evidence before it answers.

  • Docs, spreadsheets, images, recordings and transcripts
  • Hybrid search: keywords, meaning and graph links, re-ranked for accuracy
  • Self-checking: it searches again when the evidence is thin
  • Every answer cites its sources
Back to the map
Answer
10% off year two, if they sign by June 30. Approved by Maria on May 12.
Sources, ranked
Email from Maria: “Approved, 10% on renewal”96%
Contract v4, page 491%
Call notes, May 1284%
Keywords+Meaning+Graph linksRe-ranked and checked
04Know

Tools & compute

The safe building blocks agents and apps use.

Each tool is one clear action: read a report, run a calculation, update a record. Tools carry their own permissions and limits, and they run wherever the work needs them to.

  • Data, logic and action tools with clear inputs and outputs
  • Permissions and rate limits on every tool
  • MCP support, so any compatible AI app can use them
  • Runs on demand, on a schedule, in batches or as a stream
Back to the map
Tool library
crm.get_accountReadRead an account and its history
finance.marginLogicRun a margin calculation
crm.update_stageWriteMove a deal to a new stage
email.sendNeeds approvalEmail a customer
RunsOn demandScheduledBatchStreaming
MCP: any compatible AI app can use these tools, with the same permissions.
Parts 05–06

Act

How work gets done.

05Act

Agent lifecycle

Build, run and test agents like real software.

Agents are put together from tools, knowledge and rules. A lead agent can hand work to specialists, or to a person. Every agent is tested before release and scored while it runs.

  • Build: pick the role, tools, data and rules
  • Orchestrate: agents hand off to other agents or to a person
  • Evaluate: tests on every change, plus live scoring
  • Memory: agents remember past work and what worked
Back to the map
Build
RoleBookkeeperTools6DataFinance inboxRules12
Orchestrate
Lead agent
MatcherCheckerPerson
Evaluate
Correct97%Right tool99%Cost / run$0.04

An example scorecard. Every agent gets one, and it must pass before a release.

06Act

Automations

Work that starts on its own.

Not every job needs an agent. Plain automations run on a schedule, fire when something happens, or start when another system calls them.

  • Scheduled: every Monday at 7:00, or at month-end
  • Event-driven: when an invoice lands or a deal changes stage
  • API-driven: your own apps and partners can call them
  • Retries and alerts built in, so nothing fails silently
Back to the map
Scheduled
Every Monday, 7:00
Then
Send the weekly report
Ran at 7:00
Event-driven
A new invoice arrives
Then
Match it and file it
Ran 2 min ago
API-driven
Your app asks for a lead score
Then
Score and route the lead
38 calls today
Parts 07–08

Work

How your people use it.

07Work

Apps & dashboards

Tools your team actually opens.

Build internal apps and live dashboards on top of the knowledge graph, without starting from scratch. Every number links back to the records behind it.

  • No-code and low-code app building
  • Live dashboards and analytics
  • Workflow and task management
  • Scheduling and resource planning
Back to the map
Company overviewLive
Pipeline
$1.24M
up 8% this month
Cash runway
14 months
steady
Open tickets
18
down 5 today
DealStageOwner
Acme renewalContractMaria
NorthwindProposalSam
GlobexWonPriya
08Work

Teams & developers

A place for every team, and a door for engineers.

Operations, sales, finance and leadership each get assistants and views shaped for their work. Developers get APIs, an SDK, notebooks and IDE extensions to build on the same foundation.

  • Team workspaces, with assistants in Slack, email or the app
  • APIs and an SDK for your own software
  • Notebooks and a code workspace for analysts and engineers
  • IDE extensions and MCP for AI coding tools
Back to the map
In Slack
Who's on call this weekend, and did the Henderson order ship?
Sam and Priya are on call. Henderson shipped Friday. Tracking link attached.
# same knowledge, same permissions
> ask("Deals closing this month?")
7 deals · $412,000
sources: CRM (7), email (3)
> tools()
42 tools · MCP ready
APIsSDKMCPNotebooksIDE extensions
Parts 09–12

Across everything

What keeps it safe, fast and under control.

09Across everything

Security & governance

Everyone sees exactly what they should.

Access follows your org chart and your data labels, for people and agents alike. Big actions pause for a person, and risky changes are tried in a safe sandbox first.

  • Access by role, by data label and by purpose
  • Approvals and checkpoints on high-stakes actions
  • Safe sandboxes to test changes before they go live
  • Secrets kept in a vault, never inside prompts
Back to the map
Who can see what
RolePublicInternalConfidentialRestricted
Sales team
Finance team
Contractors
Bookkeeper agent
Checkpoint: pay 3 invoices over $10,000Waiting for a finance leadApprove
Testing pricing-rules-v2 in a sandbox. Live data is untouched.
10Across everything

Model gateway

Any model, behind one safe front door.

Every AI call goes through one gateway. It masks private data, picks the right model for the job, and falls back to another model if one is slow or down.

  • Commercial models: OpenAI, Anthropic, Google, xAI
  • Open-source models, your own models, or Local AI
  • Private-data masking and content filters, in and out
  • Smart caching, retries and fallbacks, with usage limits and budgets

Need it fully private? See Local AI

Back to the map
Request
Mask private data
Cacheseen it? instant answer
Router picks the model
Fast modelFrontier modelLocal modelFallback if one is down
RequestMask private dataCacheseen it? instant answerRouterFast modelFrontier modelLocal modelfallback
OpenAIAnthropicGooglexAIOpen-sourceYour ownUsage limits & budgets
11Across everything

Observability

See every step, cost and result.

Every question, agent run and automation is traced from start to finish. You can see what happened, how long it took, what it cost, and whether it was right.

  • Step-by-step traces for every run
  • Cost and speed per model, per agent, per team
  • Quality scores, with alerts when results drift
  • A full audit log of where each answer came from
Back to the map
Run #4812 · Bookkeeper agent3.2 s of work$0.03Passed
Plan the job0.4s
Search knowledge0.9s
Read QuickBooks0.3s
Match amounts1.1s
Wait for approvalOK'd
Write back0.2s
Score the run0.3s
12Across everything

Build & ship

Change it safely. Roll it back instantly.

Agents, prompts, tools and apps are versioned like code. Changes move from a sandbox to testing to production, and any version can be rolled back.

  • Versions for agents, prompts, tools and workflows
  • Separate environments: sandbox, testing, production
  • Release channels, so a change can reach one team first
  • One-click rollback to a known-good version
Back to the map
Sandbox
pricing-rules-v2
Testing
42 of 42 checks pass
Production
Finance team first
Versions
v1.7v1.8v1.9 · liveRoll back to v1.8
Deploys toManaged by usYour cloudYour hardware
How it works in practice

Two things Company OS does all day.

AWhen someone asks a question

01
Ask
Someone asks in Slack, email or the app.
02
Check
Permissions decide what they're allowed to see.
03
Find
Search pulls facts from the graph and the documents.
04
Answer
Private data is masked, and the model writes the answer.
05
Cite
Sources are attached, so anyone can check.
06
Log
The question, sources and cost are recorded.

BWhen an agent takes action

01
Trigger
An invoice lands in the finance inbox.
02
Plan
The bookkeeper agent breaks the job into steps.
03
Work
It reads the invoice, finds the order and checks the amounts.
04
Checkpoint
Anything unusual waits for a person to approve.
05
Act
The result is written back to your accounting system.
06
Learn
The outcome is scored and makes the next run better.
The engineering underneath

Built for production, not demos.

Company OS runs on the same agent architecture we use for every build. These are the parts that keep it fast, safe and reliable.

Memory
  • Working memory
  • Long-term memory
  • Memory of past runs
  • Shared team memory
Orchestration
  • Task breakdown
  • Workflow graphs
  • Retries and timeouts
  • Escalation to people
Reasoning
  • Planning
  • Self-checks
  • Confidence scores
  • Error recovery
Safety
  • Policy engine
  • Prompt-injection defense
  • Secret vault
  • Sandboxed code
Reliability
  • Replay-safe actions
  • Dead-letter queue
  • Graceful fallbacks
  • Checkpoint and resume
Cost
  • Model routing
  • Prompt caching
  • Answer caching
  • Cost tracking
Quality
  • Checks on every change
  • Nightly test runs
  • Live monitoring
  • Human feedback
Connections
  • MCP
  • Agent-to-agent protocols
  • Webhooks and APIs
  • Data lineage
Where it runs

Runs where you need it.

Managed by us

We host it, watch it and keep improving it.

In your cloud

Deployed into your own AWS, Azure or Google Cloud account.

On your hardware

Fully private, with Local AI.

Explore Local AI

Walk through the architecture
with an engineer.

Tell us what you want to dig into. We'll map it to your own tools and data.

We reply within one business day.