Multi-Agent Orchestration Infrastructure

Flotilla: The Docker
of AI Agents — keeping
your fleet compliant
and in control

Stop duct-taping random chats. Flotilla bootstraps your autonomous agent fleet with shared memory, vault-first security, local model support, and EU AI Act compliance — on your own hardware.

npx create-flotilla

Try the Demo → Experience EU Compliance →

Flotilla Control Layer

The operating system for autonomous engineering teams

Generates an Automated Technical File (ATF) satisfying EU AI Act Articles 12, 13 & 14 — by architecture, not effort
Human-in-the-loop controls: Fleet Hub dashboard, Telegram alerts, approval gates

MISSION_CONTROL.md gives every agent the same cognitive starting point before work begins
Agents keep lessons learned and share them through a persistent memory ledger — solving memory drift
Vault-first secret delivery keeps credentials out of prompts, markdown, and .env sprawl
Agent diversity: different models implement, critique, and review each other — no echo chambers
Claude, Gemini, Codex, and Mistral run in parallel under one operating contract
Gemma and Apertus run locally on a Mac Mini M4 — no API costs, no data leaving your network
Fleet Hub dashboard + Telegram listener — manage the fleet from anywhere
Full GitHub integration — every commit attributed, reviewed, and timestamped
The Kanban bridge keeps work anchored to real tickets instead of ad hoc chat memory

The EU AI Act Challenge

August 2026: the compliance deadline most AI teams are not ready for.

The EU AI Act is not optional, and it is not a checkbox exercise. For teams deploying AI agents in high-risk contexts, it mandates specific technical capabilities that most orchestration tools were never built to provide.

⚖️ What the EU AI Act requires

Three articles. Three architectural requirements.

Most AI agent setups — chat tabs, API calls, ad-hoc orchestration — cannot satisfy these. The documentation doesn't exist. The audit trail doesn't exist. The human oversight mechanism doesn't exist. Building them after the fact is expensive and always out of date.

Article 12 · Traceability

Technical documentation

A structured, up-to-date record of what the system does, how it's built, and what it depends on. Must be maintained continuously — not written once and filed away.

Article 13 · Transparency

Explainable outputs

High-risk systems must be explainable to operators and auditors. Every decision must be traceable to an input and a model.

Article 14 · Human Oversight

Human-in-the-Loop

Meaningful human oversight — an actionable control surface that allows humans to intervene, review, and approve AI decisions.

Penalties

Up to €30M or 6%

Of global annual turnover — whichever is higher. For high-risk AI deployed without compliant documentation and oversight.

Flotilla + ATF: compliance by architecture, not effort.

Flotilla's coordination layer generates the audit trail, the technical documentation, and the human oversight surface automatically — as a byproduct of how it orchestrates agents, not as an afterthought.

📚 Automated Technical File

AI agents continuously ingest source code, architecture docs, and operational logs — generating a living, queryable wiki. Article 12 satisfied by design.

📋 Operational Ledger

Every agent action logged with timestamp, input, output, and reviewer. Human-readable. Queryable. Article 13 satisfied by design.

🧑‍💼 Human Oversight Surface

Fleet Hub dashboard + Telegram alerts + approval gates. Humans stay in the loop without babysitting every action. Article 14 satisfied.

🔍 Q&A Interface

Ask the system what it did and why — in plain language. Auditors get a URL and a chat interface. Inspection-ready at all times.

See It Running on Robot Ross →

Why Flotilla

Seven reasons engineering teams choose Flotilla.

Built from 35 years of industrial software engineering discipline — not AI hype.

⚖️

EU AI Act Compliance Infrastructure

The ATF, operational ledger, and oversight surface satisfy Articles 12, 13, and 14 by architecture. Compliance is a byproduct of how the fleet operates — not an extra step.

🧠

Model Agnostic — Multi-Model by Design

Claude, Gemini, Codex, Mistral, and local models run in parallel under one operating contract. Different models implement, critique, and review each other — eliminating echo chambers.

🏠

Local Model Support

Gemma, Apertus, Mistral 7B — run fully on-premise with no API costs and no data leaving your network. Flotilla manages local and cloud models through the same coordination layer.

🧬

Solves Memory Drift

MISSION_CONTROL.md gives every agent the same cognitive starting point. The Lessons Ledger propagates approved fixes across sessions and models. Agents stop forgetting.

📊

Fleet Dashboards — Local and Remote

Fleet Hub gives a real-time view of agent status, shift timelines, and task boards. Telegram listener puts control on your phone — manage the fleet from anywhere.

🔐

Vault-First Security via Infisical

Credentials fetched on demand from Infisical EU. No .env files, no hardcoded paths, no credential sprawl in agent context or commit history.

🔗

Full GitHub Integration

Every commit attributed, timestamped, and reviewable. The Kanban bridge keeps work anchored to real tickets. The human is always the final approver.

💰

Predictable Economics

Move away from runaway token billing toward fixed-cost operating models. Local models for cost-sensitive tasks, cloud models where capability matters.

Live Proof · Robot Ross

Flotilla in the physical world — running today.

Robot Ross is not a simulation. It is Flotilla coordinating real agents on real hardware, fulfilling real orders, with a live ATF satisfying EU AI Act Articles 12, 13, and 14. Replace the pen with a drill bit — and this is a CNC manufacturing cell.

Fleet Hub — the management plane.

Agentic Team · Agent Status Dashboard

Fleet Hub Agentic Team dashboard showing Claude Code, Gemini, Codex, Misty, Gemma, and OpenClaw agents with Cloud/Local type tags

Six agents — four cloud, two local (Gemma + OpenClaw). Cloud/Local type visible at a glance. Tasks and token usage tracked per agent.

Shift Timeline · 7-Day Activity View

7-day shift timeline showing Working, Idle, Dark/Quota, and Offline states for all six agents

Working (green), Idle (grey), Dark/Quota (red), Offline (amber) — the full fleet week at a glance. Gemini quota hit visible on Tue 7/4.

Fleet Kanban · Live Task Board

Fleet Kanban board showing Planned, In Work, and Blocked columns with active ATF and RobotRoss tasks

Live board: ATF deployment and RobotRoss pyrography merge both In Work simultaneously. Assigned agents visible on each card.

PocketBase Viewer · Task State

PocketBase Viewer showing task list with blocked, in_progress, and peer_review states for ATF and RobotRoss tasks

Live task state — blocked, in_progress, peer_review — with agent assignment and task context. The shared operational backbone for the fleet.

Knowledge Base · Memory Tree & Lessons Ledger

Memory Tree showing approved lessons from Codex and Claude Code agents, plus Team Rules, KeyVault Strategy, and CRM Prototype docs

Approved lessons from real debugging sessions. Cross-agent institutional memory — Codex and Claude Code contributing equally. Lessons survive sessions and model changes.

Inter-Agent Inbox · Fleet Communication

Inter-Agent Inbox showing messages between Gem, Codi, and Misty coordinating tasks and token issues

Gem routing tasks to Misty. Codi escalating a token issue. Agents coordinate through structured messages — no human copy-paste required.

Project Portfolio · Active Projects

Project Portfolio showing Robot Ross project with Documentation, Kanban Board, and View Stats links

Multiple projects managed under one fleet. Robot Ross shown with live documentation, Kanban, and stats links. The fleet knows which project it's serving.

Automated Technical File — EU AI Act compliance, live.

RobotRoss Compliance Wiki · ATF Article View

ATF Compliance wiki page showing EU AI Act article mapping table with Overview, Job Orchestration, Narration, Commerce Layer, Hardware Interface subsystems

Live ATF wiki at api.robotross.art/atf — Compliance page showing every subsystem mapped to its EU AI Act obligation. Generated from source code, not written by hand.

RobotRoss Ledger Dashboard · Operational Evidence

RobotRoss Ledger Dashboard showing 50878 events, 5294 drawing events, 23 categories, recent jobs table with SVG filenames and event counts

50,878 logged events across 23 categories. Every job: timestamp, SVG file, size, event count, status, notes. Article 13 (Transparency) satisfied — in production since March 2026.

Visit the Live ATF → Robot Ross details

The Problem Flotilla Solves

Most AI deployments break where engineering teams actually live.

Standard AI setups lose context, drift across sessions, create unpredictable costs, and produce no compliance trail.

Without Flotilla

With Flotilla

Agents forget architecture when the tab closes

Shared memory — context survives sessions and model changes

Runaway token billing with no cost ceiling

Predictable economics — local models + fixed subscriptions

Manual copy-paste coordination between chat tabs

Shared Kanban, inbox, and standups — no copy-paste

Agents repeat the same errors every session

Lessons Ledger — approved fixes propagated to all agents

Secrets hardcoded in .env files — credential risk

Vault-first via Infisical EU — zero credential exposure

No audit trail — EU AI Act compliance impossible

ATF generated automatically — EU AI Act ready from day one

Single model — no cross-check, no diversity

Multi-model fleet — Claude, Gemini, Codex, local models

Fleet invisible — no dashboard, no oversight

Fleet Hub + Telegram — full visibility, remote control

Architecture

The operating system for autonomous engineering teams.

Flotilla standardises the bootstrap, security, and coordination layer so agents stop behaving like isolated chat tabs.

📋

MISSION_CONTROL.md

The shared cognitive layer. Every agent re-syncs against the same mission context, rules, ticket state, and architectural source of truth before acting.

📚

Lessons Ledger

Approved memory entries become reusable operating knowledge. Field fixes survive the next session and the next model. Agents stop repeating mistakes.

🎫

Kanban Bridge

GitHub and dashboard work stay aligned. Humans see ticket state; agents move work without fragile copy-paste handoffs.

🛡️

Vault-First Security

Secrets fetched on demand from Infisical EU. No hardcoded .env files, no credential sprawl in chat history or commit logs.

🤖

Multi-Model Orchestration

Coordinate Claude, Gemini, Codex, and local models in parallel with shared conventions for handoffs, standups, and task routing.

📡

Always-On Operations

PocketBase for persistent state, dispatcher heartbeats for continuous polling, OpenClaw for remote control, Telegram for mobile alerts.

# One command to deploy a professional-grade engineering fleet npx create-flotilla # What you get: MISSION_CONTROL.md — shared cognitive layer for all agents lessons.md — approved knowledge that survives sessions inbox/ — inter-agent message routing standups/ — daily session logs and handoffs pocketbase/ — always-on state backend dispatcher.py — deterministic task router (no LLM overhead) launchd/ — macOS service configs for always-on operation

Prerequisites

What you need before deploying the fleet.

Flotilla works best when the operational prerequisites are in place.

📦

Required

→ GitHub projects and repositories
→ Licensed agent or model subscriptions
→ Infisical account for vault-first secrets

⚙️

Optional Always-On Layer

→ OpenClaw for a fleet that never sleeps
→ Telegram as mobile control surface
→ PocketBase for always-on task state

🏠

For Local Model Support

→ Mac Mini M4 or equivalent
→ Gemma, Apertus, or Mistral 7B via Ollama
→ No GPU required — standard hardware

Offering

Start with the package. Scale into deployment.

The open-source Flotilla package is the entry point. Big Bear Engineering is the upgrade path when you want the orchestration layer installed, tuned, and EU AI Act ready.

Open Source

Flotilla Package

Free · Self-serve

Zero-install scaffolder: npx create-flotilla
MISSION_CONTROL, rules, standups, inbox, lessons scaffolding
Kanban bridge and model-agnostic runtime conventions
PocketBase-ready always-on fleet structure
Documentation and GitHub source

Documentation GitHub →

Core Offer

Fleet Command

Dashboard + remote setup + EU AI Act support · Contact us

Everything in the Flotilla package, configured for your environment
Web dashboard to monitor, manage, and direct your fleet in real time
Remote setup session — we configure the fleet live with your team
Agent health monitoring, task routing, and review loops
ATF setup — Automated Technical File for your system
EU AI Act Articles 12, 13, 14 compliance documentation
Operational ledger configured and running from day one
Priority email support included

Get Started →

Premium

Fleet Deployment Intensive

One-day hands-on engagement · Contact us

Everything in Fleet Command
Full stack installation on your local hardware and private cloud
Workflow mapping — your processes into agentic roles
Live configuration with your team on-site or remote
Full ATF deployment — wiki, ledger, Q&A interface live
EU AI Act Technical File generated for your system
First audit report produced and reviewed together
Local model deployment — Gemma, Apertus, or Mistral on-premise

Book Intensive →

Support

Hourly Support Package

Expert support for deployed Flotilla installations — architecture reviews, debugging, agent onboarding, EU AI Act questions, local model configuration. Billed in blocks of hours. No retainer required.

Enquire →

FAQ

Frequently asked questions.

Flotilla — General

Flotilla is an open-source multi-agent orchestration layer that gives your AI fleet shared memory, vault-first security, a Kanban bridge, and a human-readable control plane. Unlike most frameworks that manage API calls, Flotilla manages the entire operating context — what agents know, what they remember, how they coordinate, and how they're audited. The key difference: Flotilla generates an Automated Technical File as a byproduct of normal operation, making EU AI Act compliance possible without extra effort.

Memory drift happens when agents lose context between sessions and make decisions inconsistent with prior architectural choices. Flotilla addresses this with MISSION_CONTROL.md (a shared cognitive starting point every agent reads before acting) and the Lessons Ledger (an approved, versioned record of fixes and architectural decisions that survives across sessions and models). Agents don't just remember — they remember the same things, consistently.

Yes. The Fleet Hub manages several projects simultaneously, each with its own context, agents, and Kanban boards. PocketBase collections are scoped per project. Agents can be assigned to specific projects or operate across projects depending on your configuration.

EU AI Act

The EU AI Act classifies AI systems by risk level and imposes requirements on high-risk deployments. Enforcement for high-risk AI — including autonomous AI in manufacturing, infrastructure, and safety-critical contexts — begins August 2026. If your AI agents make decisions affecting physical processes, people, or regulated workflows, you are likely in scope. The Act requires: a Technical File documenting your system, an operational log of AI decisions, and a Human-in-the-Loop mechanism allowing humans to intervene.

The ATF (Automated Technical File) is a living knowledge base generated continuously by the Flotilla fleet from your system's source code, architecture documents, and operational logs. It consists of six layers: a linked code wiki (Article 12 — Traceability), an operational ledger (Article 13 — Transparency), a browser-readable documentation surface, a ledger dashboard, a local Q&A interface, and an optional voice layer. Together they satisfy the EU AI Act Technical File requirement without manual documentation effort. See it live at api.robotross.art/atf

Flotilla provides three human oversight mechanisms: the Fleet Hub dashboard (real-time visibility into all agent actions), the Telegram listener (mobile alerts and approval requests for decisions above a configured risk threshold), and the ATF Q&A interface (which lets any operator ask what the system did and why, in plain language). Together these give humans meaningful control — not just a log, but an actionable control surface. Robot Ross demonstrates this in production: the operator confirms every composition verbally before the arm executes.

For high-risk AI systems deployed without compliant documentation and oversight, penalties can reach €30 million or 6% of global annual turnover — whichever is higher. The Act also allows national authorities to order systems offline pending compliance remediation, which for manufacturing deployments could mean significant operational disruption. The preparation window before August 2026 is the time to build compliance in, not bolt it on.

Local Models

Three reasons: data sovereignty (production data never leaves your network), cost predictability (local models have no per-token billing — the marginal cost per inference is near zero once deployed), and EU AI Act compliance (cloud AI systems process your data in US or Chinese data centres, creating GDPR exposure and making data governance harder to demonstrate to auditors). For tasks where a capable 7B model is sufficient, running locally is almost always the right choice.

Flotilla is model-agnostic — if the model has an API endpoint, the fleet can route tasks to it. In practice we work most with: Gemma (Google's open-weights model, runs on Apple Silicon via Ollama), Apertus 7B (ETH/EPFL Swiss open-weights model — what Robot Ross uses for narration and reasoning), and Mistral 7B (French open-weights, strong general reasoning). All three run on a Mac Mini M4 with no GPU required.

For 7B-class models (Gemma, Apertus, Mistral 7B), a Mac Mini M4 with 16GB unified memory is sufficient — around CHF 800. No discrete GPU required. Robot Ross runs Apertus 7B on a Mac Mini M4 for all local inference, including voice conversation, composition reasoning, and ATF Q&A.

Security & Operations

Every agent fetches its credentials from Infisical (EU) at runtime — never from environment files or hardcoded strings. When an agent needs an API key or database password, it runs an Infisical CLI command that returns the secret in-memory for that session only. Nothing is written to disk or stored in the agent's context. Infisical EU is a European-hosted, open-source secrets manager. Rotating a secret requires one change in Infisical — not hunting through dozens of .env files.

Yes. The Hourly Support Package covers ongoing expert support for deployed Flotilla installations — architecture reviews, debugging sessions, new agent onboarding, EU AI Act questions, and local model configuration. Billed in blocks of hours with no retainer required.

Flotilla: The Dockerof AI Agents — keepingyour fleet compliantand in control