I
One company loop
Every department follows the same grammar: sense, decide, perform, verify, record, learn. Commercial work is simply the first route through it.
Company Architecture · Living Blueprint · 02 Sep 2026
The Company
Assembly Line
A product-agnostic operating spine for an AI-native company. It turns goals and signals into governed work, verifiable outcomes, and reusable learning—continuously, across commercial operations, product, delivery, content, and software. People set direction and handle exceptions; the line keeps moving inside those boundaries, even when everyone is out of office.
- Scope
- Whole company
- Current wedge
- Commercial line
- Human role
- Govern · review
- Operating mode
- Continuous
- Portability
- Any product
- Architecture
- Harness-agnostic
01
The spine, not another workflow
The assembly line is the durable operating model of the company. Individual flows, agents, models, channels, and products can change around it.
II
Autonomy by policy
Agents may act only inside explicit goals, permissions, budgets, and gates. Anything outside that envelope becomes a structured escalation.
III
Memory outside agents
Agents are replaceable workers, not the company brain. Business truth, run state, evidence, and operating rules live in shared systems.
IV
Always on, never unbounded
Pre-authorised work continues off-hours. High-impact work waits safely for review; absence never silently expands authority.
The unit of architecture is not the lead. It is the work order: a durable request to move a company object from one state to another, under policy, with evidence. A lead, product issue, support case, content brief, renewal risk, and pricing experiment are different objects travelling through the same machinery.
02
The company spine
Direction flows down; evidence and outcomes flow back. The middle loop can run repeatedly without losing state or inventing authority.
Direction & authority
PurposeGoalsBudgetsPoliciesApproval rights
Shared company memory
Business recordsRun checkpointsArtifacts & evidencePolicies & skillsEvents & metrics
Every transition is identifiable · resumable · observable · attributable · replay-safe
The control plane owns the lifecycle: admission, routing, waiting, retries, approvals, timeouts, and completion. The execution fabric performs bounded steps. Systems of record own truth. This separation is what lets a new assistant, model, tool, or department join without becoming a second company brain.
03
The fundamental company model
The assembly line needs a small, stable vocabulary that describes any product company before it describes a specific campaign or tool.
What the company is trying to change, by when, within which limits.
What exists, what it can do, how it performs, and what is planned.
Who receives which value, at what price, with what message and evidence.
Company
kernel
kernel
The shared operating truth
The relationships among value, market, work, authority, and outcomes—not a collection of disconnected automations.
IdentityOwnershipStateHistoryEvidence
Leads, customers, users, partners, segments, behaviour, and intent.
Work orders, stations, agents, tools, budgets, queues, and commitments.
Revenue, adoption, quality, incidents, cost, feedback, and knowledge gained.
The universal objects
Company object
Anything whose state matters: product, offer, person, account, opportunity, issue, release, case, asset, invoice, experiment.
Signal
A change worth noticing: a message, event, threshold, schedule, intent signal, error, customer action, or human request.
Work order
The durable contract for desired change: objective, context, authority, success criteria, owner, lifecycle, and evidence.
Station
A reusable capability that transforms work: research, qualify, compose, test, approve, send, deploy, reconcile, or analyse.
Artifact
A durable output or input: brief, email, proposal, report, patch, test result, recording, screenshot, contract, or dataset.
Event
An immutable statement that something happened. Events wake work, advance state, feed observation, and preserve history.
Gate & escalation
A decision boundary. A gate checks policy; an escalation packages ambiguity or risk for the right human to resolve.
Outcome
The measured effect after work leaves the line. Completion is not success until its business or operational result is observed.
04
One line, many routes
Departments are named routes through shared machinery. Select a route to see how the same six-stage grammar serves different parts of the company.
Commercial line
The first active route: from market signal to qualified relationship and measurable revenue learning.
Objects: account, person, lead, opportunity, offer, activity, conversation, outcome. Key invariant: every external contact is attributable to an approved offer, audience policy, and work order.
Product & offer line
Turns market and customer evidence into product choices, packaged value, and experiments.
Objects: product, capability, problem, offer, price, audience, experiment, roadmap item. Key invariant: product and offer decisions retain their evidence and the outcome that later confirmed or rejected them.
Software factory
Converts issues and observations into tested, reviewed, released, and monitored improvements.
Objects: issue, incident, repository, change, test, review, release, service, metric. Key invariant: code generation is not completion; verified behaviour in the live product closes the order.
Customer line
Coordinates onboarding, support, success, renewal, and feedback without fragmenting the relationship record.
Objects: customer, subscription, onboarding plan, case, request, health signal, renewal, feedback. Key invariant: every hand-off preserves context and gives the customer one coherent company, not a chain of agents.
05
The company stations
A station is a durable capability, not an org-chart box. The same station can serve several routes and be staffed by an agent, software, a service, a human, or a combination.
STgovern
Strategy & portfolio
Maintains goals, priorities, budgets, product bets, and the authority under which the rest of the line operates.
- Consumes
- outcomes, constraints, opportunities
- Produces
- goals, policies, allocation decisions
POcreate
Product & offer
Defines the value the company can deliver and turns it into coherent capabilities, packages, prices, proof, and roadmap choices.
- Consumes
- strategy, evidence, customer needs
- Produces
- products, offers, experiments
CMoperate
Commercial
Discovers and develops relationships from signal through research, qualification, engagement, opportunity, and revenue outcome.
- Consumes
- offers, market signals, policy
- Produces
- relationships, pipeline, insight
CBamplify
Content & brand
Converts company knowledge into reusable narratives, assets, publishing, distribution, and observed audience response.
- Consumes
- insights, proof, product truth
- Produces
- content, reach, demand signals
CXserve
Customer & delivery
Coordinates onboarding, fulfilment, support, success, retention, and the return of customer evidence to the whole company.
- Consumes
- commitments, context, requests
- Produces
- value delivered, health, feedback
SWmaintain
Software factory
Builds and repairs the digital machinery: triage, implementation, tests, review, release, operational validation, and recovery.
- Consumes
- issues, roadmap, incidents
- Produces
- verified changes, releases, fixes
OPcontrol
Finance, risk & operations
Controls money, obligations, vendors, compliance, entitlements, and the high-impact boundaries agents may not cross alone.
- Consumes
- commitments, spend, exceptions
- Produces
- controls, approvals, reconciliations
OBobserve
Observation & intelligence
Measures business outcomes and machinery health, detects anomalies, explains change, and emits new signals into the line.
- Consumes
- events, costs, quality, outcomes
- Produces
- metrics, alerts, diagnoses, evidence
HRauthority
Human review & governance
Sets policy, evaluates consequential decisions, resolves ambiguity, handles escalations, and changes the autonomy envelope.
- Consumes
- briefed decisions and exceptions
- Produces
- judgement, approval, correction
06
The work order is the universal job
Agents may reason freely inside a step; the company never relies on an agent conversation to remember what the job is, what it may do, or whether it finished.
WORK ORDER · WO-01842DURABLE CONTRACT
- Objective
- The desired business or operational state change, expressed independently of the worker or tool.
- Subject & context
- Company objects, source signal, relevant history, and links to authoritative records.
- Owner & route
- Accountable route, current station, assigned executor, priority, and deadline.
- Authority
- Allowed actions, data access, spending limit, communication policy, and approval gates.
- Success
- Acceptance criteria, verification method, expected artifact, and target outcome.
- Recovery
- Current checkpoint, attempt history, idempotency key, timeout, and compensation path.
- Evidence
- Decisions, tool results, artifacts, costs, confidence, review, and observed outcome.
Restartability comes from the work order and its checkpoint, not from a chat transcript. A worker can stop, a model can change, and a tool can fail. The control plane resumes at the last confirmed transition and uses idempotency to avoid repeating an external side effect.
07
Humans govern by exception
The objective is not zero humans. It is zero unnecessary human coordination: people receive decisions, evidence, and exceptions—not a pile of agent activity to reconstruct.
Autonomous
Read-only research, enrichment, classification, analysis, internal summaries, draft creation, testing, and routine observation.
continues off-hours
Bounded action
Known, reversible actions inside a policy: low-risk record updates, scheduled follow-ups, internal task creation, approved content preparation.
continues within limits
Review gate
External communication while trust is being earned, pricing changes, customer commitments, sensitive data use, release approval, or unusual spend.
waits safely
Human authority
Legal commitments, banking, ownership changes, exceptional refunds, employment actions, policy changes, and expansion of agent permissions.
never delegated silently
An escalation contains
What happenedOne-sentence state and why the line cannot proceed.
Decision neededTwo or three explicit options, recommendation, and consequence.
EvidenceSource records, artifacts, confidence, policy, cost, and deadline.
When the office is empty, work inside the pre-authorised envelope continues. Gates remain gates; they queue with context and deadlines. Timeouts can reroute or alert, but they cannot convert absence into approval.
08
The software architecture underneath the line
A connected system view of how signals become durable work, how agents execute it, and which system owns each record. This describes software responsibilities and connections—not hosting or deployment.
Human surfaceSlack + HermesRequests, approvals, escalations, status, and read-only answers.
IngressWebhooks + schedulesProvider events, deadlines, recurring work, and monitors.
Business signalsRecord + product eventsTwenty lifecycle changes and PilotX or customer-system events.
AuthorityHuman operatorsGoals, corrections, interventions, approvals, and policy decisions.
Durable control planeActivepiecesNormalises signals, creates work orders, owns run state, routes stations, waits, retries, resumes, requests approval, verifies, and completes.
Versioned recordGitSoftware, policies, skills, prompts, schemas, tests, and reviewed automation definitions.
Business recordTwentyProducts, offers, companies, people, leads, opportunities, customers, relationships, and activities.
Agent executorAgent ZeroPlans the bounded station assignment, delegates, uses tools, and returns structured status and evidence.
Reasoning + coding fabricHarness adaptersCodex, Claude, Gas Town–style coordination, and future local or API models.
External capabilitiesConnectorsResend, Gojiberry, LinkedIn, browser/computer use, PilotX, GitHub, and future services.
Artifact recordR2Evidence, screenshots, journey captures, reports, exports, recordings, and generated assets.
Company feedback loopObservation + metric layerCanonical events, decisions, cost, latency, quality, delivery, incidents, and business outcomes become new signals.
work, data, and controloutcome and feedback
company_idproduct_idoffer_idsubject_idwork_order_idrun_idartifact_id / event_id
One work order across the systems
01Signal arrives and receives an event id
02Twenty resolves the business subject and offer
03Activepieces creates the work order and checkpoint
04Git supplies approved policy, skill, and test version
05Agent Zero dispatches the best harness and tools
06R2 stores outputs and evidence by artifact id
07Connectors report verified external effects
08Business state, outcome, metrics, and next event commit
Record ownership and write rules
| Record family | Authoritative owner | Written by | Referenced elsewhere as | What it must not become |
|---|---|---|---|---|
| Business object | Twenty | Controlled record adapters after validation | company_id, offer_id, subject_id | A flow-local copy that drifts from the customer or commercial truth. |
| Work order & run state | Activepieces | Control-plane transitions only | work_order_id, run_id, current station | The CRM, artifact store, or long-term knowledge base. |
| Artifact & evidence | R2 | Stations and connectors through the artifact contract | artifact_id, checksum, type, retention, source | An anonymous file path with no subject, provenance, or work order. |
| Operating definition | Git | Reviewed commits and approved automation exports | repository, version, commit, policy or skill id | Mutable runtime state or a queue of unfinished work. |
| Company event | Canonical event contract | Control plane, record adapters, connectors, and observers | event_id, type, source, time, subject, work order | An untyped provider payload that every downstream flow interprets differently. |
| Delivery status | Channel provider + company event | Resend, Gojiberry, LinkedIn, or future channel adapter | provider id + work order id + recipient identity | “Sent” inferred from an attempted call with no provider confirmation. |
The event and metric layer is a logical requirement even if V1 begins with Activepieces run history and Twenty activities. As volume grows, it can gain a dedicated store without changing the event contract or the rest of the line. Git still should not be used as a live database or restart mechanism.
09
Hermes is a doorway, not a second control plane
Slack, email, a web interface, and future assistants are company entry points. They should all see the same truth and create the same durable work.
Entry points
- Slack — conversation, approvals, escalations, status
- Hermes — concierge, interpreter, and company navigator
- Email — inbound requests, replies, and notifications
- Webhooks — signed events from external systems
- Schedules — recurring work, deadlines, and monitoring
- Interfaces — dashboards and specialised operator views
Hermes interaction contract
01
Recognise the person, company context, intent, and required authority.
02
Answer read-only questions from shared records, with sources and freshness.
03
For durable or consequential work, create or update a work order in the control plane.
04
Return status, approval requests, results, and evidence to the same conversation.
05
Keep no private operational truth that other assistants or routes cannot access.
Two paths
- Fast path — safe, read-only, immediate answers with no durable side effect.
- Durable path — work that waits, changes state, sends, spends, coordinates, or must survive interruption.
- Specialist path — Hermes delegates a bounded step to a specialist but retains the shared work-order identity.
More assistants do not require more architectures. A Grok-style Slack assistant, Hermes, a product copilot, and a software agent can all join as surfaces or workers. They authenticate, read shared context, and hand durable work to the same control plane. This prevents fragmented queues, private memories, and conflicting authority.
10
Stable capabilities, replaceable tools
The architecture names responsibilities first. The current stack is one implementation of those contracts, so individual tools can be replaced without redrawing the company.
| Stable capability | Responsibility | Current implementation | Replacement boundary |
|---|---|---|---|
| Business system of record | Owns products, offers, people, organisations, relationships, leads, opportunities, customers, activities, and business state. | Twenty | Replace only through a record adapter that preserves identity, history, ownership, and event semantics. |
| Durable control plane | Owns triggers, work-order progression, waits, retries, schedules, approvals, timeouts, and external side-effect coordination. | ActivepiecesIts internal storage is run state—not the authoritative business record. | Flows call stable station and record contracts; they do not embed a model or vendor as company logic. |
| Agent executor | Decomposes bounded work, delegates steps, uses tools, returns structured results, and reports evidence. | Agent Zero | Executor receives a work order and returns status, artifacts, events, and escalation—never private company state. |
| Harness & model fabric | Supplies specialised reasoning and coding capacity behind a common execution boundary. | Codex · Claude · other harnessesGas Town–style coordination or API/local models can join per task. | Adapters let a route select capability, cost, policy, and quality without depending on one CLI or model provider. |
| Artifact & evidence store | Holds documents, datasets, generated assets, recordings, exports, screenshots, and large execution evidence. | R2 | Work orders keep immutable references, metadata, checksums, retention, and access policy—not storage-specific paths. |
| Versioned operating memory | Owns software, agent definitions, skills, policies, schemas, prompts, tests, and reviewed operational changes. | Git | Git is versioned truth and change control; it is not the live queue, CRM, or substitute for resumable run state. |
| Communication connectors | Deliver and receive messages while reporting provider status and the resulting relationship events. | Resend · Gojiberry · LinkedIn | A channel adapter normalises send, receive, identity, consent, delivery, reply, failure, and idempotency. |
| Human interaction surface | Presents conversations, approvals, escalations, interventions, summaries, and operational status. | Slack · Hermes | Any surface must use shared identity and work orders; it cannot become an unobserved side channel. |
| Event ingress & egress | Receives signed webhooks, normalises external changes, suppresses duplicates, and emits trusted company events. | Activepieces webhooksA dedicated event gateway can be introduced when volume or governance requires it. | Canonical events isolate the company loop from provider-specific payloads. |
| Observation & intelligence | Connects business outcomes, run health, cost, quality, incidents, and agent behaviour into actionable signals. | Shared event & metric layer | Observers consume canonical events and never become the only location where operational truth exists. |
The control plane persists coordination; systems of record persist truth. Activepieces can remember that a flow is waiting and resume it. Twenty remembers the commercial relationship. R2 remembers the artifact. Git remembers the reviewed operating definition. None should impersonate all four.
11
V1 — the evidence-led PilotX spine
The first concrete production line turns a qualified brand into verified evidence, a commercial conversation, measured customer value, and reusable learning. Select a station to inspect its contract.
AI-operated commercial assembly line
Qualified brand → verified evidence → commercial conversation → measured value → reusable learning
Human supervision rail
Review evidence and claims, first external contact, pilot activation, commercial commitments, and unsafe or uncertain actions.
Evidence decision branch
Actionable gap: continue to conversion. No viable gap: archive or nurture as a valid completed result.
1. Source — discover and qualify
Fill the line with brands that are worth testing and have a plausible buyer.
Input
- Ideal-customer profile
- Territory and category constraints
- Research pool or inbound signal
Work
- Discover and deduplicate brands
- Confirm fit and relevant commerce signals
- Identify likely decision-makers
Output
- Qualified Brand Pack
- Company and buyer record
- Recommended journeys and fit rationale
Pass gate
- Brand fits the agreed profile
- Store and buyer are reachable
- A controlled test can produce evidence
Escalate when
- Ownership or identity is ambiguous
- Lawful-contact basis is unclear
- Fit signals conflict materially
V1 is one offer-specific spine, not the whole assembly line. Its source may be outbound today, but any feeder that produces a qualified brand and a testable hypothesis can enter Station 1. Its learning output returns to qualification, test design, evidence standards, messaging, product, and future strategy selection.
12
V2 — many strategies, one factory
Dria’s suggestions expand how demand and proof enter the company. They do not require disconnected pipelines: each strategy emits a canonical signal, enters the strategy portfolio, and is routed into the right offer spine.
Strategy feeders · examples, not a closed list
all feeders emit attributed demand signals
Strategy portfolio & experiment ledgerHypothesis · audience · source · asset · offer · spend · variant · consent · conversion event · owner
→
Offer and spine routerResolve fit and intent, select the relevant promise and production line, or hold the signal for nurture and learning.
route by offer, evidence path, and required fulfilment
Source → Test → Prove → Convert → Deliver → Learn.
Outcome: qualified retained-revenue customerAcquire → Capture → Analyse → Publish/API → Subscribe → Learn.
Outcome: paid audit, data product, or standalone toolRecruit → Qualify → Enable → Co-sell → Deliver → Expand.
Outcome: partner-sourced customers and reusable proofShared factory floorCompany kernel · work orders · Activepieces · Agent Zero · harness adapters · Twenty · R2 · Git · connectors · observation · human governance
Architectural translation of the meeting ideas
A Convert-station experiment. Both variants use the same Evidence Snapshot and produce attributed conversion events.
An approved benchmark artifact added at Prove, then cited at Convert. Claims remain traceable to evidence and assumptions.
A feeder when it captures intent; a distinct offer spine if users pay for the result itself.
Reusable artifacts become distribution assets; engagement returns as attributed signals to Source.
A feeder for the V1 offer at first; if enablement and white-label fulfilment differ materially, promote it to its own spine.
A Learn-stage outcome and proof asset that feeds product credibility, AEO, customer workflows, and future conversion.
Car-company rule: a new acquisition strategy is another road into the plant; a materially different offer is another assembly line inside the plant. Neither requires a second operating system. New ideas can always be added as feeders, tested against outcomes, and either scaled, revised, or retired.
13
Observation closes the line
An AI-native company does not merely perform more actions. It can explain what happened, connect work to outcomes, repair its machinery, and improve under review.
Transitions, attempts, latency, cost, errors, tool use, approvals, and artifacts.
Replies, revenue, activation, retention, satisfaction, quality, and delivery.
Corrections, overrides, review reasons, confidence, and escalation patterns.
Observe
→ improvework · product · policy · tools
→ improvework · product · policy · tools
Change sequencing, station boundaries, retries, gates, context, or success criteria.
Update roadmap evidence, audience, positioning, proof, price, and qualification.
Turn defects, regressions, toil, and missing capabilities into verified change orders.
Every action emits enough evidence to answer: what triggered this, why was it allowed, who or what acted, what changed, what did it cost, did it meet the standard, and what happened afterward? Observation then creates new signals—some for business routes, some for the software factory, and some for human governance.
Learning proposes; governance changes. Agents may identify a better prompt, policy, skill, flow, model, or threshold, then test and submit the change. They do not silently rewrite their own authority or the company’s operating rules.
14
Portable by changing the product pack
To operate another product, keep the company spine and replace the product-specific truth. The same routes and stations can then be configured, measured, and improved for a new context.
Reusable company kernel
- Work-order contract and lifecycle
- Control-plane and executor boundaries
- Identity, permissions, gates, and escalations
- Stations, events, artifacts, and observation
- Human review and change-governance model
+configure
Product pack
- Product, capabilities, constraints, and roadmap
- Offers, audiences, prices, promises, and proof
- Knowledge, language, policies, and regulated boundaries
- Accounts, channels, connectors, and available actions
- Goals, metrics, budgets, quality bars, and review thresholds
Product ACommercial route first
Product BCustomer route first
Internal platformSoftware route first
The assembly line is therefore reusable without pretending every company behaves identically. The machinery stays stable; the product pack, enabled routes, policies, integrations, and autonomy envelope express the differences.
15
How the company grows into the blueprint
Build outward from a real route, while preserving the whole-company contracts from the beginning.
- Prove the commercial route.Connect signal capture, business records, qualification, offer selection, engagement, replies, follow-up, human review, and outcome measurement through durable work orders.
- Add the shared observation layer.Make every step explainable and connect delivery metrics to business outcomes, costs, reviews, and failure patterns.
- Feed improvement into the software factory.Let incidents, workflow defects, product feedback, and operator friction become triaged, testable, reviewable engineering work.
- Open additional routes.Introduce content, product-and-offer, customer, finance, and operational stations without forking identity, memory, or authority.
- Add assistants as shared doorways.Bring in Hermes and other specialists through the entry-point contract so the experience expands without multiplying control planes.
- Package the kernel for another product.Separate reusable company machinery from product truth, then instantiate a new product pack, autonomy policy, and route priorities.
16
The rules that keep it one company
These invariants matter more than any named tool. If one is broken, the system may be automated but it is no longer a coherent assembly line.
- One authoritative owner for each kind of truth.Business record, run state, artifact, policy, and metric have explicit homes and stable identities.
- No consequential side effect without a work order.External messages, commitments, spend, record changes, and releases are attributable, governed, and replay-safe.
- Agents are replaceable; company memory is not.No essential state lives only in a model context, local scratchpad, or assistant conversation.
- Every pause can resume.Waiting for a person, event, provider, or retry does not strand the job or require someone to reconstruct it.
- Every completion is verified.A sent request, generated patch, or updated record is an action; success is a checked result and, eventually, an observed outcome.
- Authority narrows automatically and expands deliberately.Uncertainty, missing context, increased impact, or policy conflict causes a gate or escalation—never improvised permission.
- Humans receive decisions, not archaeology.Reviews and escalations arrive with context, options, recommendation, evidence, and consequences.
- Learning is versioned and reviewable.The line may propose and test changes to itself, but policies, skills, flows, and permissions change through explicit governance.
The intended end state: the company keeps sensing, deciding, acting, verifying, and learning while the team is away—without losing its memory, crossing its authority boundaries, or becoming unable to explain what it did.