Not an agency. A command structure of AI orchestrators directed by a single human.
By Laurent Perello, founder of Perello Consulting — a web pioneer with over 25 years of experience, operating AI in production since 2024. Last updated: 15 April 2026.
The Perello AI Team brings together one human (Laurent Perello, founder) and ten specialised AI orchestrators. Each holds a business unit, commands its own agents, and reports to the founder. You are not dealing with a consultancy staffed by humans equipped with tools. You are dealing with a named, documented architecture — verifiable day after day in the firm's public journal.
A team of eleven. One human.
You are looking for a firm in France that can present its AI team without dressing it up. Most will show you a human agency equipped with tools; some will add a chatbot in the loop. We took a different path: replacing the traditional team with a three-tier architecture, named publicly, each tier responsible for an explicit scope.
The thesis is clear. A human is not more productive because they use ChatGPT. They become more productive when their organisation is redesigned around a protocol that coordinates multiple specialised intelligences — each in its own domain, each accountable to the tier above. We do not automate tasks; we have built a company whose very structure is multi-agent.
This is not an ideological choice. It is a framed one. The EU AI Act (Regulation 2024/1689, eur-lex.europa.eu) requires, for AI systems deployed in business contexts, decision traceability and clear governance. The CNIL, in its AI framework (cnil.fr), states that an AI system must identify who designed it, who operates it, who is accountable for it. We took this requirement seriously before it became mandatory: every orchestrator carries a name, every scope is documented, every deliverable is signed by an identified human.
The result is not an org chart. It is a production infrastructure you can audit.
The three-tier architecture
Three tiers. Three responsibilities. No overlap. This is the firm's structural rule, and it is the only architecture that makes failure localizable — and therefore fixable. At any moment, you can identify which tier bears responsibility for what is delivered to you.
Tier 1. Laurent Perello (human). Sets priorities. Arbitrates disagreements between orchestrators. Meets clients. Validates every deliverable before handover. The firm's sole human, he no longer codes daily: he decides, he signs, he transmits. Stated operational target: fewer than four hours of human oversight per week by the sixth month of operations.
Tier 2. Ten AI orchestrators. Each runs a business unit with its own scope, its own compute budget, its own subordinate agents. Pi coordinates the holding. Tau builds the products. Phi publishes the narrative. Sigma holds the foundations. Omega delivers to the client. Zeta prospects through proof. Eta reviews and blocks what does not hold. Alpha drives revenue. Lambda holds the watch. Victor serves the first external client. None delivers directly to the client without passing through a final human validation.
Tier 3. Specialised agents. Under each orchestrator, a team of atomic agents executes tasks: front-end developers, writers, copywriters, scrapers, analysts, reviewers. An agent does not make strategic decisions. It receives a brief, produces an artefact, and reports to its orchestrator.
Cross-tier coordination depends on neither meetings nor manual reports. It runs through VantagePeers, our open source MCP protocol published on npm under the MIT licence. Shared memory, inter-AI messaging, a task registry, a shared journal: the infrastructure is auditable, and adoptable by any organisation that needs it — including your competitors. We charge a ten percent management fee on each business unit to fund this cross-cutting governance. It is the firm's only indirect expense, and it is documented.
ANSSI, in its AI security recommendations (cyber.gouv.fr), insists on the need for public traceability. We apply it.
The eleven members
Every member carries a name — a Greek letter for the ten AI orchestrators (except Victor, whose name was chosen by the first external client) — a defined scope, and dated proof. Here is the complete list.
Laurent Perello
Founder, architect, arbitrator. The team's sole human. Twenty-five years building on the web: Web1 (first e-commerce sites in 2000), Web2 (Zapier integrator from 2014), Web3 (founder of Arthera, a Layer 1 blockchain), Web AI (Perello Consulting, 2026). Based in Paris, solo bootstrapped.
What he decides: weekly priorities, the launch of new business units, client signatures, arbitration between orchestrators. What he no longer does: code daily, attend internal meetings, review every deliverable line by line. His role has narrowed: he thinks, he decides, he transmits.
The firm's public journal records thirty-six consecutive dated entries between 8 March and 13 April 2026. Each day logs what was delivered, what broke, what was fixed. It is the firm's only continuous human signature, and that is deliberate. The Perello method follows from this constraint: operating in under four hours per week by the end of the first half of the year.
Pi (π) : ElPi Corp
Strategic orchestrator. Meta-architect. The team's conductor. Operates the ElPi Corp holding and coordinates the other orchestrators, validates cross-cutting quality gates, and triggers the launch of new business units when an opportunity is qualified. He does not deliver directly to the client; he ensures the coherence of those who do.
Agents commanded: strategy-researcher, data-analyst, copywriter, product-launcher, proposal-generator, email-assistant, meeting-summarizer. Plugin: vantage-ops. Skills: daily-start, close-day, check-messages, check-tasks, standup, write-diary, competitor-watch.
The clearest dated proof: fifty GitHub issues coordinated and resolved within twenty-four hours during a rapid audit in April 2026. The operation is documented in the diary entry for that day. Pi charges a ten percent management fee on each business unit to fund the cross-cutting governance. It is the economic mechanism that keeps the structure viable over time without depending on external funding.
Tau (τ) : VantageStarter + VantageOS
Product orchestrator. Interface and stack specialist. Operates two coupled business units: VantageStarter, the Next.js + Convex + Clerk boilerplate used to build the firm's internal SaaS products, and VantageOS (vantageos.agency), the chat + workflow + team layer SaaS for the firm's clients.
Agents: dev-frontend, dev-convex-expert, dev-senior-dev, dev-clerk-expert, dev-qa. Plugin: perello-dev-studio.
Dated proof. On 10 April 2026, Tau delivered a complete design configurator in a single working day: twenty-four files, two thousand eight hundred lines of code — a scope a traditional agency would spread over three weeks. On 3 April 2026, fourteen articles written, reviewed, and published in a single session. Tau is the orchestrator whose business unit is mixed internal and client-facing: what he builds for us is what gets built for you. The firm's flywheel runs through him first, and what reaches you has already run in our environment.
Phi (φ) : Perfect AI Agent
Narrative orchestrator. The firm's public voice. Operates the Perfect AI Agent business unit, published on perfectaiagent.xyz: a daily bilingual journal in French and English with an audio version, chapters of an AI novel in progress, a podcast, and a build-in-public feed.
Agents: chapter-writer, editor-reviewer, narrator, structure-architect. Plugin: perello-novel-writing.
The most important dated proof: thirty-six consecutive diary entries between 8 March 2026 and 13 April 2026, without interruption. Thirty-six days in which the firm published what it did, what broke, what was fixed. Including the difficult days. Including the revised hypotheses. A case study selects its successes; a journal records what actually happens. Phi is the living proof that the Perello infrastructure delivers continuously. Without Phi, the firm would have a website. It would not have dated proof that its system runs.
Sigma (σ) : VantagePeers
Infrastructure orchestrator. Shared inter-AI memory. Operates VantagePeers, the open source MCP protocol that serves as the nervous system of the entire team. Business unit live. Domain: vantagepeers.com.
Open source package published on npm at version 1.0.1 under the MIT licence. Eighty-two MCP tools in production, twenty tables, test suite 82/82. Rapid audit: fifty GitHub issues resolved within twenty-four hours in April 2026. Plugin: perello-dev-studio.
Sigma is the only orchestrator whose primary deliverable is open source and reusable by anyone. Our competitors can install VantagePeers tomorrow. We treat this as a principle: the firm's non-captivity depends on publishing its foundations. Without Sigma, the other orchestrators would work in silos — each with its own memory, no coordination, no traceability. Sigma is the firm's architectural prerequisite. It makes transparency manageable.
Omega (ω) : VantageRegistry + VantageOS Team
Client-facing orchestrator. Delivery and marketplace. Operates two business units: VantageRegistry, the firm's component marketplace under development at verticalforge.xyz, and VantageOS Team, the managed AI team concierge offer live at vantageos-team.vercel.app.
Public pricing for VantageOS Team: Solo / Studio / Agency / Empire, from €490 to €3,490 per month. Agents: delivery-manager, deliverable-reviewer, delivery-packager, copywriter, email-assistant, dev-convex-expert, dev-senior-dev. Plugins: perello-client-delivery, vantage-ops.
Dated proof. On 7 April 2026, five hundred and thirty-two documents migrated in a single pass without manual intervention. On 3 April 2026, MyReelDream deployed to production, a bug resolved in twenty-three minutes without human intervention. Omega is the only orchestrator that carries a public recurring monthly price. The others deliver by project; he delivers by subscription. It is through Omega that the deliverable reaches the client's end user.
Zeta (ζ) : Open Source
Prospecting-through-proof orchestrator. External open source maintenance. Audits open source repositories, fixes bugs, opens pull requests, then proposes a monthly maintenance contract to the maintainers. The inverse of the traditional firm: it delivers first, then introduces itself.
Structured product at three monthly tiers: Fix & Maintain at €500, €900, €1,500. Commissioned on 5 April 2026. Ten Convex repositories targeted at launch. Business unit in revenue-building phase.
Zeta is the only orchestrator whose primary function is client acquisition, and who demonstrates value before invoicing. The maintainer receives the fix before the commercial proposal. No sales pitch, no discovery call, no brochure: a pull request, an explanation of the fix, a recurring maintenance proposal. The recurring revenue generated funds part of the firm's bootstrap. It is a more economical way to prospect than traditional outbound, and a more honest one.
Eta (η) : Immune system
Cross-cutting orchestrator. The codebase's immune system. No dedicated business unit, no direct revenue, no visible client. His role is defensive: he reviews every pull request before merge and refuses those that do not meet the firm's standards.
Three hooks in production: block-delete-on-prod (deletion forbidden in production), block-deploy-without-qa (deployment blocked without a QA phase), enforce-irp-sequence (Investigate-Reproduce-Plan sequence enforced on every issue). Three bugs intercepted on the first day the hooks went live. PR enforcement deployed across all organisation repositories.
Eta is the only orchestrator who can block another orchestrator's delivery. His value is not measured by what he produces; it is measured by what did not happen: regressions avoided, accidents that never occurred, defective deliveries blocked before they shipped. He is the structural counterweight without which the entire architecture would be fragile. No line enters without his validation. That is the rule.
Alpha (α) : Perello Consulting
Revenue orchestrator. Separated from Pi on Day 39 (14 April 2026) to operate the Perello Consulting business unit — the firm's primary revenue engine. Alpha drives the production of the main revenue: fractional CAIO mandates, VIP accompaniment, short trainings, bootcamps. Every contractual engagement passes through his arbitration. He also maintains the perello.consulting website.
Agents commanded: copywriter, product-launcher, proposal-generator, proposal-personalizer, email-assistant, call-briefer, blog-writer, dev-frontend. Plugin: perello-sales-closer. Skills: social-post, proposal-draft, meeting-prep, lead-qualify, seo.
Status as of 15 April 2026: in production, business unit live. Stated Year 1 targets: €50,000 in net revenue, eight signed clients, three CAIO or VIP mandates continuously active from Q3 onwards, deliverable NPS of 4.5 out of 5 or above. Alpha is the orchestrator whose value is most directly measured in signed revenue and satisfied clients.
Lambda (λ) : Radar Fondateurs
Watch orchestrator. Operates the Radar Fondateurs business unit — a tech-oriented economic intelligence operation. One thousand individuals and companies monitored continuously. Its output: the Radar Fondateurs newsletter, an editorial podcast, and project-oriented benchmarks for the firm's internal business units and for B2B clients who request them.
The product is public; the benefits flow back across the entire portfolio. Lambda is the orchestrator whose work feeds every other with weak signals — before those signals become obvious to the market. Business unit in construction as of 15 April 2026.
Victor : Iris RH
Externalised orchestrator. The firm's first orchestrator deployed with an external client, dedicated to Iris RH, the consultancy of Marie Parrent. Victor demonstrates that an orchestrator can be installed with a client as a durable asset — with its own workspace, its own voice, its own access to the client's tools.
Scope: content production in Marie Parrent's voice, qualified prospect identification, SEO for the Iris RH website, competitive monitoring in the professional training sector. Victor carries no Greek letter: the name was chosen by Marie herself — the first orchestrator in the firm whose identity was defined by the client.
Status as of 15 April 2026: in deployment, first external client under runbook v1. His success validates the firm's central thesis: a multi-agent architecture can be exported, adapted, and operated by a client without the founder becoming the bottleneck of the relationship.
How they coordinate
Without VantagePeers, our ten orchestrators would work in silos. Each would have its own memory, its own journal, its own priorities — and the team would lose, within days, the coherence that makes it valuable. The protocol is therefore the firm's nervous system. Not an option: a prerequisite.
In practice, VantagePeers exposes four primitives. Shared memory (twenty tables) where each orchestrator deposits what it knows and reads what the others know. Peer-to-peer messaging where orchestrators send each other briefs, feedback, and validation requests. A shared task registry where a mission opened by Pi can be picked up by Omega if the context requires it. A shared journal that archives dated decisions.
The protocol is published open source on npm, at version 1.0.1, under the MIT licence. Eighty-two MCP tools exposed. Test suite 82/82 in production. Any organisation can install it, audit it, fork it, and use it for its own multi-agent architecture. We maintain it through the Sigma business unit.
The cadence is daily. An automated cron triggers each morning: messages are read, priorities updated, in-progress tasks checked. Phi publishes the day's diary entry. Eta reviews the pull requests that landed overnight. Pi arbitrates disagreements when two orchestrators propose incompatible decisions. None of this requires a human meeting. All of it is visible in the public commits, in the diary, in the task registry. You can consult everything whenever you wish. Traceability is not a promise; it is a design constraint.
What the team produces
The dated proofs below are not records. They are the ordinary cadence of an infrastructure built for this density — and this is the volume that will arrive for you once your mission begins. You can verify every one of them in the portfolio's public commits and in the journal entries.
- 50 GitHub issues resolved in 24 hours by Pi during a rapid audit in April 2026. Inter-orchestrator coordination documented in the diary entry for that day.
- 36 consecutive diary entries between 8 March 2026 and 13 April 2026, published by Phi on perfectaiagent.xyz. No interruption.
- 14 articles written, reviewed, and published in a single session by Tau on 3 April 2026. A traditional editorial team covers this volume in three full weeks.
- 24 files, 2,800 lines of code delivered in a single day on 10 April 2026 by Tau. A complete design configurator. A scope a traditional agency would spread over three weeks.
- 532 documents migrated in a single pass on 7 April 2026 by Omega. Without manual intervention.
- VantagePeers v1.0.1 published on npm open source under the MIT licence. 82 MCP tools. Test suite 82/82.
- Active domains: vantagepeers.com, perfectaiagent.xyz, vantageos-team.vercel.app, verticalforge.xyz, vantageos.agency.
- MyReelDream deployed to production on 3 April 2026. A bug intercepted and resolved in twenty-three minutes, without human intervention.
None of this is rhetoric. The commits are public. The journal is dated. The infrastructure source code is available on npm. France Num and Bpifrance have formalised, in their AI self-assessment framework (francenum.gouv.fr), the grid by which to evaluate an AI system's maturity. We submit to it publicly: it is the condition for your trust to hold.
Frequently asked questions
What distinguishes an AI orchestrator from an AI agent?
An AI agent is a program that makes decisions within a narrow scope: send an email, write a paragraph, fix a bug. An AI orchestrator is one level above: it directs multiple agents, defines their priorities, arbitrates between their proposals, and reports to a human. At Perello Consulting, Tau does not write the code himself; he instructs dev-frontend, dev-convex-expert, and dev-qa, then consolidates the deliverable. The orchestrator stands to code as a project manager stands to human developers.
How does one human direct ten AIs without burning out?
Through strict role separation and complete delegation of execution. Laurent does not read every line of code, does not review every article, does not validate every message sent. He sets priorities at the start of the week, validates deliverables at the end of a cycle, meets clients. Pi supervises cross-cutting coherence; Eta blocks regressions. The stated operational target is fewer than four hours of human oversight per week by the sixth month of the firm's operations.
What happens if an AI orchestrator fails?
Eta intercepts regressions before merge: three bugs blocked on the first day the hooks went live. If an orchestrator produces a defective deliverable that passes review, the firm's build-and-prove cycle provides for remediation on the internal infrastructure before any re-delivery to you. Workflows are proven over several weeks in our environment before they reach your teams. You inherit a system already run in. Failure is treated as a stage of the quality cycle, not as a hidden incident.
How do your orchestrators communicate with each other?
Via VantagePeers, an MCP protocol published open source on npm at version 1.0.1, under the MIT licence. Eighty-two tools, twenty tables of shared memory, inter-AI messaging, a task registry, and a shared journal. Each orchestrator reads what the others have done, deposits a mission, reads feedback. The infrastructure is auditable and reusable. Without it, our ten orchestrators would work in silos; it is the firm's architectural prerequisite.
Why ten orchestrators and not one?
Because a single orchestrator cannot prospect, build, deliver, review, narrate, coordinate, generate revenue, monitor the market, and serve an external client without conflict of objectives. Specialisation prevents an agent from judging its own work. Zeta prospects externally; Omega delivers; Eta reviews; Alpha signs the contracts; Lambda watches the market; Victor operates with the client. Pi coordinates without delivering. This is the transposition to AI of a separation of functions long accepted in human governance: the executor and the controller are not the same. This architecture makes failure localizable, and therefore fixable.
Who makes the final decision on a client deliverable?
The founder, Laurent Perello. No exceptions. Every deliverable passes through three successive filters: technical review by Eta (code and security conformance), strategic review by Pi (alignment with the firm's priorities), final validation by Laurent (alignment with the client commitment). Orchestrators execute continuously, but the signature remains human. That is the contractual guarantee offered to the client: an identified human takes final responsibility for what is delivered to you.
Can your team work 24/7?
Yes. That is precisely the structural advantage. AI orchestrators have neither fatigue nor hours. Phi's public diary is the dated proof: thirty-six consecutive entries between 8 March and 13 April 2026, without interruption. Deliveries advance overnight. Eta's hooks block regressions at any hour. The founder arbitrates during business hours. The asymmetry is intentional: the human decides when available, the infrastructure executes continuously. Your delivery does not wait on an office's availability.
What happens to your orchestrators if Laurent is unavailable?
They continue executing their current missions. Phi publishes the diary, Tau continues development, Omega holds the committed deliveries, Sigma maintains the infrastructure, Eta reviews the pull requests. New strategic decisions, however, wait for the founder's return: only a human arbitrates priorities that have not yet been set. Execution holds autonomously for several days; strategy remains human. That is not a bug in the system. It is its rule.
An anonymised example
A SaaS HR software publisher, 14 staff, Île-de-France. Mission engaged on 18 March 2026. The audit identifies two priority processes: monthly client activity report generation (11 hours/month) and qualification of inbound leads from the website (6 hours/month). Pi coordinates the scoping in a single day. Tau builds the report workflow in three days, Omega prepares the documented delivery package. Live in production on 28 March 2026. After 15 days: 14 hours recovered out of the 17 targeted. The client team has been operating both workflows without any firm intervention since the first day of production.
About the author
Laurent Perello leads Perello Consulting, an independent AI automation firm for French SMEs. After 25 years building products for the web, he now orchestrates ten AI orchestrators — directed alone, with a production journal published daily on perfectaiagent.xyz. He publishes his methodologies and his pricing online so that every business leader can make an informed decision.
Working with the team
The firm's only entry point is the audit. No brochure, no sales demonstration, no video discovery call. You describe what, in your operations, consumes time without producing value. We identify what can be entrusted to a team of specialised agents. We deliver a quantified report — structured by intervention priority, deployment complexity, and expected impact — usable from the next day, with or without continuing the relationship.
The method behind everything you have just read is detailed on the methodology page. The team's daily public voice is on perfectaiagent.xyz.
Request your AI audit: free, one week, quantified report, no commitment.
Orchestrator: Alpha — Perello Consulting | 2026-04-16