AI operating system
See blueprint
Blueprint for AI-assisted operations

Turn context
into action.

An architecture to automate pre-sales and pilots by connecting versioned knowledge, specialized agents, executable skills, and business systems.

Explore the systemSee a full process
Versioned memory Connected systems Human control
One source of truthExecutable processesControlled integrationCumulative learning
The starting point

The problem is not the model.
It is the context.

A generic agent can draft text. An agent with rules, memory, tools, and procedures can operate consistently.

Before
  • Scattered information
  • Follow-up by memory
  • Repeated manual work
  • Individual knowledge
With the system
  • Structured context
  • Visible follow-ups
  • Automatable flows
  • Shared learning
Architecture

A harness that tells the agent what to know, what to do, and how far to go.

The system combines a structured knowledge base with executable procedures and secure connections to the tools where the work happens.

Sources
CRM Calendar Take-notes Web documentation APIs
MCPAPI
Orchestration
Context-aware agentreason · propose · execute · verify
RulesSkillsGuardrailsTools
ReadWrite
Persistent memory
context-repo/main
├── operating-rules.md├── processes/├── products/├── skills/└── accounts/<organization>/
Versionableevery change has historyPortablenot tied to a sessionAuditabledecisions with evidence
01
Declarative knowledge

What we know

Internal processes, product briefs, concepts, variables, success criteria, and deep technical documentation.

Process guidesExecutive briefsTechnical references
02
Operational knowledge

How we act

Skills that turn recurring procedures into executable instructions, with steps, tools, and checks.

WorkflowsScripts and APIsDiagnostics and recipes
03
Living context

What is happening

Status, scope, decisions, meetings, follow-ups, and outcomes for each process, versioned over time.

Current stateFollow-upsHistory and evidence
End-to-end processes

From a commercial request to a traceable process.

Automation is not a single action: it is a chain that interprets, verifies, asks for approval, writes to the right systems, and keeps context.

01

Structured request

The agent interprets the commercial requirement, separates facts from assumptions, and detects missing information.

02

Triage and verification

It checks that a use case, a product, and a real technical need exist. It looks for duplicates before creating anything.

03

Draft for approval

It proposes process type, pipeline, description, owners, and fields; a person confirms before anything is written.

04

CRM + repository

It creates the ticket and, in the same flow, opens or updates the versioned context with a uniform structure.

05

Session and propagation

It turns the meeting into a note, status, follow-ups, and CRM progress without recapturing everything by hand.

06

Follow-through to close

It queries follow-ups and stages, assigns the right ownership, and keeps decisions and outcomes traceable.

Technical session

Pre-sales

The commercial role keeps global progress. The technical role owns agreements, answers, and engineering tasks.

RequestTriageSessionClose
Controlled trial

Pilot / POC

The technical role owns organization and follow-up: real scope, success criteria, checklist, dates, risks, and results.

DiscoveryPre-kickoffActiveResults
Calendar + take-notes

The meeting stops being an event. It becomes state.

The agent crosses calendar, attendees, and note sources; it checks that real content exists and propagates agreements once.

Calendarevent · attendees · date
Notessummary · transcript · agreements
Agentclassify · deduplicate · propose
Repository CRM note Message
One review block. Zero writes before approval.
Specialized skills

From written knowledge to executable knowledge.

A skill packages the when, the how, the allowed tools, known failure modes, and how to verify the result.

SKILL.mdoperational
trigger: pilot requestinputs: context, scope, datetools: CRM, calendar, repositoryworkflow: validate → propose → approveverify: ticket + folder + linkguardrail: do not write without confirmation
01

Triggered by intentThe agent recognizes when to apply the procedure.

02

Shrinks the decision spaceIt orders steps, sources, and exceptions before acting.

03

Limits the toolsIt declares least privilege and approval gates.

04

Accumulates experienceReal errors become gotchas and recipes.

Process skills

They orchestrate work across people, systems, and the repository.

  • Request → ticket
  • Meeting → update
  • Calendar → daily plan
  • Scope → proposal

Product skills

They turn technical documentation into safe API operations.

  • Configure and diagnose
  • Create flows and integrations
  • Run validations
  • Messaging and signing

Infrastructure skills

They prepare technical changes with review and controlled deployment.

  • Domain configuration
  • Versioned changes
  • Pull requests
  • Deployment checklist
API operating pattern

Discover. Test. Connect. Verify.

Product skills combine documentation, scripts, and diagnostics to turn a complex API into a repeatable, safe sequence.

1Read fresh stateGET
2Test the actionTEST
3Apply the full documentPOST
4Verify the resultGET
5Record the learningGIT
Webhook over pollingTest before integratingToken only in the environmentPost-write verification
MCP + APIs

The agent does not replace the tools. It orchestrates them.

Connectors standardize access to business systems. Direct APIs cover deep product operations. The repository keeps the context that joins both worlds.

CRM

Tickets, pipelines, notes, owners, duplicate search, and follow-up.

MCP · controlled read and write

Calendar

Meeting discovery, attendees, acceptance, and time context.

MCP · structured query

Take-notes

Summaries, transcripts, and agreements turned into actionable context.

MCP · content retrieval

Web documentation

On-demand lookup of official references and behavior validation.

Web / API lab · research

Product APIs

Configuration, diagnosis, tests, and repeatable operations with scripts.

Direct API · specialized execution

Team messaging

Progress, alerts, and questions about missing information in the working channel.

Connector · proactive communication
M
MCP connectors

Business tools

  • Managed authentication
  • Operations with a known schema
  • Contextual read and write
  • Permissions scoped by skill
+
/
Direct API

Product operations

  • Reproducible scripts
  • Deep technical diagnosis
  • Tests and transformations
  • Ephemeral credentials
Result

One continuous flow

Context → decision → action → evidence

Governance

Automate without losing control.

Useful autonomy does not mean unlimited access. The system defines what it can read, what it can prepare, and what needs human confirmation.

Agent analyzessources + rules + context
Drafts a changechanges and visible effects
Human approvescorrects or authorizes
Agent executeswrites + verifies + records

Human before write

Tickets, external notes, proposals, and high-impact changes go through a preview and approval.

Secrets stay out of the repository

Tokens and credentials live in environment variables; never in documents, conversations, or commits.

Facts separated from inferences

The agent confirms scope, validates sources, and asks under ambiguity instead of filling gaps with assumptions.

Minimum relevant data

It keeps commercial and technical context, not sensitive data about end users or private evidence.

Every action leaves a trace

Changes are dated, versioned, and tied to a source, a decision, or an owner.

Bounded blast radius

Each skill declares which tools it may use and verifies state before and after a mutation.

Operational impact

Savings show up where friction used to live.

It is not only about generating text faster. It reduces search, context switching, repeated typing, lost agreements, and dependence on individual knowledge.

01Retrieval

Less searching

Status, decisions, and documentation stop living across chats, notes, and individual memory.

02Operations

Less double entry

The same conversation feeds the repository, the CRM, and the follow-up channel.

03Execution

Less reinvention

Skills reuse API procedures, diagnostics, and guardrails already learned.

04Follow-up

Fewer forgotten processes

Every missing piece of information becomes a follow-up, and every ticket has associated context.

05Knowledge

Better onboarding

A new person or agent can understand the process by reading the same versioned source.

06Governance

Better control

Automation prepares and executes; sensitive decisions keep a human gate.

Learning
ExecuteVerifyCorrectVersion
Memory that improves

Each operation strengthens the next one.

Corrections are not lost in a conversation. They become rules, calibrations, references, edge cases, and new verification steps.

  • History of decisions and changes
  • Follow-ups as future questions
  • Gotchas built into skills
  • Calibration against real outcomes
Implementation path

Layered evolution, not a leap of faith.

Autonomy is built once memory, procedures, and permissions are already trustworthy.

Level 1Foundation

Structured memory

Rules, processes, templates, and knowledge live in Git.

Level 2In use

Repeatable assistance

Skills guide frequent tasks and reduce variability.

Level 3Evolving

Connected orchestration

MCP and APIs link CRM, calendar, notes, and product.

Level 4Goal

Bounded autonomy

Loops review follow-ups, ask questions, and report with guardrails.

A reusable pattern

Knowledge that is not only queried.
Knowledge that operates.

Repository as memory. Agents as operators. Skills as procedures. MCP and APIs as connections. People as owners of the decisions.

Context+ Agents+ Connections= Augmented operations