OctaFlow

Project Rapido · v1 · draft for board review

An org chart of agents, and the documents that make them work

This document defines the organizational architecture for the two agentic companies we run on the Project Rapido stack — GEB (Global Economic Bridge) and Octaflow — as an org chart of generic, reusable roles, and it specifies, per role, which organizational documents must feed that agent's system prompt and why.

An agent in this system is not a chat persona. It is a role whose behaviour is the sum of four Paperclip artifacts (AGENTS.md, SOUL.md, HEARTBEAT.md, TOOLS) plus the source documents those artifacts point at. The artifacts are the shape of the role; the organizational documents are the substance. A role with a good SOUL.md but no company documentation to reference will produce confident, well-worded, wrong work. The point of this document is to make the substance explicit so Aman and Rajan know exactly what to write and hand over.

Scope

Role definitions, hierarchy, the document catalogue, the role×document matrix, and a compulsory/good-to-have classification. Out of scope: the per-role AGENTS/SOUL/HEARTBEAT/TOOLS themselves.

A role is its artifacts plus the documents they point at

01 / 08 · Purpose

Five rules the architecture follows

Design principles

One role catalogue serves both companies; documents feed prompts rather than being copied into them; the org spans two planes; human oversight is a role; and every company starts small and grows by need.

1 · Generic roles, adapted per company One role catalogue serves both companies; GEB and Octaflow adapt each role rather than inventing bespoke titles. This mirrors the Paperclip companies/ library — the same handful of roles (CEO, CTO, QA, Reviewer, Release, Research) recur across 16 companies and 440+ agents; specialization is configuration, not new archetypes.

2 · Documents feed prompts; prompts don't embed them A system prompt references the vision, the architecture, the conventions, and it does not copy them. The documents live in the seam repo (read scope per role) so they change once and every agent inherits the change. Same "config per product, not code per product" rule the GEB Shared Runtime follows.

3 · Two planes, one org Governance roles live in Paperclip (plan, approve, gate); execution roles live in Fusion (build, run, ship). The org chart spans both; the seam is where a governed spec becomes an executed task.

4 · Human oversight is a role Following the Headcount Zero "founder + kill switch" model, the board sits at the top of the chart as the human principals who own the source documents, set strategy/budget, and hold the stop authority. Agents propose; humans decide the things agents are not allowed to decide.

5 · Start small, grow by need Every document is classified compulsory or good-to-have. A company can launch on the compulsory document set and author the good-to-have ones as volume justifies. The org is staffed on the same principle — a lean compulsory role spine first.

Reference, don't copy — one source, many prompts

02 / 08 · Design principles

The hierarchy

The org chart

Solid lines are reporting/delegation. The governance spine is Board → CEO → (COO, CTO). The quality function runs in two tiers on opposite planes — the plane that builds never judges: QA / Quality Auditor (Tier 1) does deterministic step-boundary checks on the execution side, while the Tier 2 Evaluation Owner owns the async LLM-as-judge on the governance side and reports to the CEO. Security / Compliance is compulsory for both companies (Octaflow because, as the skills producer, it is the larger supply-chain surface). Roles below the dotted "good-to-have" line — including the new Compiler / Antenna — are absorbed by their parent until volume justifies a dedicated agent.

Board → CEO → COO / CTO, over an execution & assurance band

03 / 08 · The org chart

Generic role, two adaptations

Role catalogue

Each role is generic; the GEB and Octaflow columns show how it is adapted. "Plane" is where the agent primarily lives.

The org spans two planes, joined by the seam

#	Role	Mandate (generic)	Plane	GEB adaptation	Octaflow adaptation
1	CEO (lead)	Own goals/budget proposals, delegate, run the blocking quality gate, sign delivery	Paperclip	Approves each reviewed deliverable before it reaches a buyer	Approves each House deliverable / shipped skill
2	COO	Throughput, coordination, budget tracking, keep the board moving	Paperclip	Keeps runs flowing across products; watches per-run cost	Keeps the 6 Houses / 4-tier ladder moving; watches build cost
3	CTO	Own the technical system + the Architect gate; runtime, skills, seam, evals	both	Owns the GEB Shared Runtime + per-product config	Owns the build stack, repo standards, CI, the seam
4	Chief of Staff	Route incoming work, groom backlog, reduce CEO polling	Paperclip	Triages buyer submissions into runs	Triages House requests / marketplace intake
5	Skills Architect / Librarian	Own the skills library: Penelope sync, agentshield scan, `skills.lock`	both	Curates the skills each product resolves to	Curates Marketplace / House skills; skill production
6	Workflow Designer	Design the runtime/workflow patterns the engineers execute	Fusion	Maintains the 9-stage decisioning pattern	Designs House-specific pipelines
7	Runtime / Lead Engineer	Execute the work: run the runtime (GEB) or build features (Octaflow)	Fusion	GEB Runtime Engineer — executes a run per mandate	Lead Engineer — builds skills/products
8	Build Engineers	Backend / frontend / fullstack build capacity	Fusion	Rarely needed (research house)	Backend/Frontend/Fullstack per House at scale
9	Research Analyst	Gather and classify evidence for a decision or a build	Fusion	Core — source material, seller/spec verification	Market/competitor research for a House
10	QA / Quality Auditor (Tier 1)	Tier 1 — deterministic checks at the step boundary; does not judge (the plane that builds never judges)	Fusion	Deterministic gate compliance on each run	Deterministic gate compliance on each skill build
11	Code Reviewer	Review build work for correctness before merge	Fusion	N/A (little code)	Reviews skill/product code pre-merge
12	Security / Compliance Auditor	agentshield, sanctioned-jurisdiction, evidence-policy enforcement	both	Compulsory — sanctions + evidence integrity	Compulsory — producer-side gate (skills supply-chain surface)
13	Release / CI Engineer	Merge, ship, deploy; own the pre/post-merge gates	Fusion	Publishes the reviewed deliverable	Ships skills to the Marketplace / deploys
14	Design Reviewer / Creative	Brand, voice, UI, marketing output	Fusion	Light — deliverable formatting/voice	Octaflow creative/marketing Houses
15	Tier 2 Evaluation Owner	Own the LLM-as-judge, its calibration and drift (Tier 2, async); reports to the CEO — an auditor never reports to the builder (may be CEO-held at launch)	Paperclip	Judges reviewed deliverables against the rubric / evidence policy	Judges House deliverables / shipped-skill outputs
16	Compiler / Antenna	Standing curation of the catalogue: keep it current, watch the field, compile the derived registers (D14) (good-to-have; CTO-absorbed at launch)	both	Keeps the catalogue + living registers current	Keeps the catalogue + living registers current
—	Board / Principals (human)	Strategy, budget, hiring approval, source documents, kill switch	above both	Aman/Rajan/David	Aman/Rajan/David

The minimum viable agent gate

Before any role on the chart is instantiated as an agent, it passes one question: does this seat's work run repeatedly, and can its output be scored? No to either, and the seat stays a deterministic step or a human job for now — the box remains on the chart as intent, not a build commitment. This keeps the compulsory spine small and honest, and it is the bridge to the Composition Layer.

04 / 08 · Role catalogue

The source documents

Organizational document catalogue

These are the source documents that nourish the agents' system prompts; they live in the seam repo and are read-only to most roles. Each entry now carries a mutability class — the catalogue should not treat all documents as the same kind of thing.

Mutability classes & the change-through-a-gate rule

S · Static — board-authored, read-only. G · Governed-mutable — a living register agents propose changes to and humans ratify, through a change gate. D · Derived — compiled from feeds/registers, never hand-authored (the Compiler / Antenna maintains it).

A document that many agents inherit is changed through a gate, never silently — every G change is proposed, ratified, and logged in D15. Most of D1–D13 are class S; three entries split or join below.

Sixteen documents, held in the seam, referenced by prompts

Key	Class	Document	What it contains	Why an agent needs it
D1	S	Vision & Charter	Why the company exists, mission, mandate, north-star, non-goals	Without it, agents optimize locally and drift from intent; it is the tie-breaker for every judgment call
D2	S	Product & Use-case Catalogue	The products/use cases and their per-item config (GEB: product cards; Octaflow: 6 Houses + 4-tier ladder)	Tells the agent what it is producing and for whom; the runtime resolves specifics from here
D3	S	System / Software Architecture	The two-plane stack, the seam, northcheck deployment, runtime design, external interfaces	Agents that touch the system must know its shape or they build against the wrong seam
D4	S	Development Conventions	Repo layout, `PROMPT.md` spec, branch/merge rules, the Agent Companies standard, naming	Keeps execution agents' output mergeable and reviewable; the "house style" of building
D5	S	Skills Catalogue & Production Method	The skill file architecture (SKILL/GATES/SECURITY/VOICE/modules), the Changes Register, Penelope as source of truth	Skills/runtime agents must know how a skill is structured, versioned, and imported safely
D6a	S	Quality Gates — deterministic	Deterministic gate definitions (Architect / step-boundary gates); changes with the build stack	The Tier 1 pass/fail acceptance criteria enforced deterministically at the step boundary
D6b	G	Judge Rubrics	The LLM-as-judge scoring rubrics and reference sets; every scored run grows them	The Tier 2 evaluation criteria the judge scores against (owned by the Tier 2 Evaluation Owner)
D7	S	Security, Compliance & Evidence Policy	SECURITY manifests, agentshield rules, sanctioned jurisdictions, evidence classes (Verified/Claimed/Not-found), trust boundaries	Hard, deterministic constraints; the difference between a defensible deliverable and a liability
D8	S	Brand & Voice Guide	VOICE_CARD, tone, output modes, formatting standards	Anything customer-facing must sound like the company; prevents off-brand deliverables
D9	S	Budget & Economics Model	Model lanes/costs, per-company budgets, cost-per-run targets ("do the math")	COO/CEO must trade throughput against spend; sets the cheap-execute/strong-review split
D10	S	Roadmap & OKRs	Goals, priorities, short/long term plan, sequencing	Turns the vision into what to do this cycle; the CEO/COO planning input
D11	S	Glossary & Domain Terminology	Canonical terms (GEB commodity terms; Octaflow: Lark Base, AnyCross, Ask Penelope, House names)	Prevents term drift and mis-resolution across agents and models
D12	S	Data & Integrations Spec	APIs (Paperclip, Fusion), Langfuse/OTel, webhooks, external interfaces	Agents that call or measure the system need the contract, endpoints, and auth model
D13	S	Governance & Escalation Policy	Approval gates, human-in-the-loop points, roles/permissions, the kill switch	Defines what agents may decide vs. must escalate; the guardrail on autonomy
D14	D	Living Registers Index	The accumulating assets: failure records, the loop catalogue, routing and cost records	They feed prompts exactly as D1–D13 do; the record the system learns from (maintained by the Compiler)
D15	G	Changes Register (documents)	The audited log of every proposed and ratified change to a mutable (G) document	Makes the change-through-a-gate rule auditable: who changed what, under whose sign-off
D16	S	Seam Contract Spec	The interface contract between the two planes (governance ↔ execution); owned by the CTO	An unowned interface is where organisations fail silently; agents crossing the seam bind to it

05 / 08 · Document catalogue

Which documents feed which prompt

Role × document matrix

Which documents feed each role's system prompt. R = required (the role is unsafe/ineffective without it) · H = helpful (improves quality, not load-bearing) · blank = not needed.

Reference resolution

An R or H resolves to the named sections a role needs, not the whole document — 13+ documents referenced whole is more than some seats can hold attention over, so the unit of reference is the section. This changes nothing in the grid; it changes what a cell means when authored.

Paperclip finding (from the Paperclip corpus review)

Paperclip loads whole files into an agent's context (a multi-file bundle exposes the other files but does not auto-load them) and has no native section-level loading. So section-scoped references are realized by authoring (splitting a document into section files, or assembling the section at prompt-composition time); that mechanism belongs to the Composition Layer, not this catalogue.

Role \ Doc	D1	D2	D3	D4	D5	D6a	D6b	D7	D8	D9	D10	D11	D12	D13	D14	D15	D16
CEO	R	R	H		H	R	R	R	H	R	R	H		R	H	R	H
COO	R	R	H	H	H	H		H		R	R	H	H	R	H
CTO	R	R	R	R	R	R	H	R		H	H	R	R	R	H	H	R
Chief of Staff	R	R				H		H		H	R	R		R
Skills Architect	H	R	R	R	R	R	H	R	H			R	H	H			H
Workflow Designer	H	R	R	R	R	R	H	H			H	R	H				H
Runtime / Lead Eng	H	R	R	R	R	R	H	R	H			R	R	H			R
Build Engineers		R	R	R	H	R		H				R	R				H
Research Analyst	H	R				R		R	H			R	H
QA / Quality Auditor (T1)		R	H	H	R	R		R	H			R
Code Reviewer		H	R	R	H	R		R				R	H
Security / Compliance	H	R	R	R	R	R		R				R	R	R			H
Release / CI Eng		H	R	R		R		H		H		H	R	H			H
Design Reviewer	H	R		H		R		R	R			R
Tier 2 Evaluation Owner	H	R				H	R	R				H		H	R	R
Compiler / Antenna	H	H			H							H			R	R

How to read a row

The CEO's row says the CEO system prompt must reference the Vision (D1), the use-case catalogue (D2), the gate + judge definitions and evidence policy (D6a/D6b, D7), the budget (D9), the roadmap (D10) and the governance policy (D13). It does not need dev conventions (D4) or the data/integration contract (D12) — those are execution concerns it delegates.

How to read a column

Any document with many R's is a keystone. D2 (use-cases), D6a (gates) and D7 (security/evidence) are required by nearly every role and should be authored first; D3 (architecture) and D11 (glossary) are the next tier.

From the Organisation Layer review

The D6a/D6b split, the D14/D15/D16 columns and the two new rows (Tier 2 Evaluation Owner, Compiler / Antenna) come from the review. Ratified cells: D6b/D2/D7/D14/D15 = R for the Tier 2 Evaluation Owner and D14/D15 = R (D1/D11 = H) for the Compiler, plus the five corrected cells (COO·D5, Security·D4, Design Reviewer·D7, Release/CI·D11, and Chief of Staff·D3 confirmed blank).

06 / 08 · Role × document matrix

Which documents to write first

Compulsory vs. good-to-have documents

Which of the source documents (D1–D16) a company must have before it can run, and which can be authored as it scales. Compulsory = a governed, gated, safe cycle cannot run without it. Good-to-have = it raises quality, consistency, or economics; a first cycle can run without it. (Separate axis from the S/G/D classes: a document can be Static and good-to-have, or Governed-mutable and compulsory.)

Nine compulsory documents, eight authored as you scale

Compulsory docs

Good-to-have docs

Total (D1–D16)

Compulsory documents

Doc	Why it cannot be skipped
D1 · Vision & Charter	The tie-breaker for every judgment call; without it agents optimize locally and drift from intent.
D2 · Product & Use-case Catalogue	Tells every agent what it produces and for whom; the runtime resolves specifics from here. Required by nearly every role.
D3 · System / Software Architecture	Agents that touch the system build against the wrong seam without it.
D5 · Skills Catalogue & Production Method	The runtime resolves skills per product; safe import (agentshield, `skills.lock`) needs the method.
D6a · Quality Gates (deterministic)	The Tier 1 acceptance criteria for every output; no Architect/CEO/QA gate can run without it.
D7 · Security, Compliance & Evidence Policy	Hard, deterministic constraints; the line between a defensible deliverable and a liability. Hardest limit for GEB.
D12 · Data & Integrations Spec	The seam, the bridge and the eval rail can't be wired without endpoints, auth, and the OTel/Langfuse contract.
D13 · Governance & Escalation Policy	Defines what agents may decide vs. must escalate — and holds the kill switch.
D16 · Seam Contract Spec	The interface the governed spec crosses to reach execution; an unowned/undefined seam is where the two-plane cycle fails silently.

Good-to-have documents (author as you scale)

Doc	Why it can wait / when it becomes compulsory
D4 · Development Conventions	Keeps build output mergeable; a research cycle runs without it. Compulsory for Octaflow (a build shop).
D8 · Brand & Voice Guide	Polishes customer-facing output. Compulsory once shipping customer-facing deliverables / Houses.
D9 · Budget & Economics Model	A first cycle runs without a formal cost model. High-priority for GEB / the COO — cost-per-run is the Rapido thesis.
D10 · Roadmap & OKRs	You can run one workflow without a full roadmap; it turns vision into this-cycle priorities as work scales.
D11 · Glossary & Domain Terminology	Prevents term drift — the highest-R good-to-have (a keystone by density); a cycle runs while it is being written.
D6b · Judge Rubrics	The Tier 2 judge criteria; a first cycle runs on the deterministic gates (D6a) until the judge is stood up and calibrated.
D14 · Living Registers Index	The system's accumulating record (failures, loops, routing, cost); it grows as runs accumulate, so nothing to hold at launch.
D15 · Changes Register	The audited log of changes to mutable (G) documents; it arrives with the first G document, not before.

Minimal viable document set, per company

GEB (research/decisioning house): the 9 compulsory documents, with D7 authored to full depth first (sanctioned jurisdictions + evidence classes) and D9 + D11 (commodity terms) as the first good-to-have adds — evidence integrity and run economics are the product.
Octaflow (build/skills shop): the 9 compulsory documents, with D4 (conventions) and D8 (voice) promoted to compulsory because it ships built, customer-facing House deliverables; D5 (skills method) carries extra weight as the Marketplace / skill-production backbone.

07 / 08 · Compulsory / good-to-have

What this document is built on

Sources consulted

The role models and platform behind the catalogue, the OctaFlow internal briefs that supply the substance, and the prior Project Rapido artifacts this generalises.

External references — role models & platform

Paperclip example companies
github.com/paperclipai/companies — 16 companies & rosters; the recurring roles behind §4 (Superpowers, AgentSys, Aeon, TÂCHES, RedOak)↗ Paperclip platform
github.com/paperclipai/paperclip — the agent-as-a-role model and the four artifacts (AGENTS/SOUL/HEARTBEAT/TOOLS)↗ Headcount Zero / Zero-Employee Company book
github.com/AnthonyDavidAdams/zero-employee-company-book — "founder + kill switch"; the Board / Principals layer (§2.4)↗ Paperclip Company Playbook (Aron Prins)
github.com/aronprins/paperclip-company-playbook — the blocking CEO / Architect quality gate and the maturity rule (D6)↗

OctaFlow internal briefs & architecture

Project Rapido operating manual — context-docs/project-rapido-operating-manual.md. The two-plane stack (Paperclip governance → Fusion execution), the seam and the Agent Companies standard, Langfuse-over-OTel measurement, models-per-layer. Basis for §2.3 and D3 / D12 / D13.
High-level architecture proposal (v2) — project-rapido/03-architecture/high-level-architecture-proposal.md. Northcheck deployment, the seam as a shared repo, the Penelope skill-sync plugin. Basis for D3 and D5.
GEB build plan — context-docs/geb_build_plan.html. The GEB delivery engine, the Architect gate, and the per-product rubric/scoring. Informed the CTO's Architect-gate mandate and the GEB adaptations in §4.
Octaflow build plan — context-docs/octaflow_build_plan.html. The six Houses, the 4-tier ladder, and the Gates (slop / Architect / security). Informed the Octaflow adaptations in §4 and D2.
GEB Shared Runtime pattern — context-docs/Penelope-Shared-Runtime.workflow.json and project-rapido/03-architecture/guides/simulated-geb-workflow-guide.md. The product-agnostic decision runtime and its 9-stage pattern; source of the "config per product, not code per product" principle (§2.2) and the Workflow Designer / Runtime Engineer roles.
Skill-production methodology — CLAUDE.md, context-docs/Skill_Builder_Briefing_for_Fede_v0_4.md, context-docs/OctaFlow_Engineering_Onboarding_Brief_v2.md. The skill file architecture (SKILL / GATES / SECURITY / VOICE / modules) and the Changes Register. Basis for D5 and D6.

Review & external evidence

Organisation Layer review (Aman) — the source of the three seat decisions, the S/G/D mutability classes, the D6 split + D14/D15/D16, the five matrix corrections, and the minimum-viable-agent gate.
MetaGPT & ChatDev — multi-agent ablations showing that removing role definitions from system prompts was the largest single performance hit either team measured (external proof for shape + substance).
Measurement Rail work — the two-tier quality model (Tier 1 deterministic / Tier 2 LLM-as-judge), the ratified routing rules, and Langfuse-over-OTel as the eval instrument (Tier 2, D6b, D14 lean on it).
Antenna design — the Compiler seat (standing catalogue curation / derived registers).

Prior Project Rapido artifacts

CEO & CTO agent artifact sets (GEB, in geb-sim/) — the already-compiled four-artifact examples this document generalises into the role × document mapping and the compile order.

08 / 08 · Sources consulted