Multi-agent system for a company department: coordinating several agents under a single point of supervision
A company department — a team, a division, not a local authority — never handles just one kind of task: letters, case files, follow-ups and summaries come one after another, often in tools that do not talk to each other. A multi-agent system brings several specialised AI agents into line — each assigned to a precise task — under overall steering and supervision. Hosted in France, with local inference or an isolated resource, it replaces no decision: each agent produces a deliverable that can be checked, approved by your teams.
Updated on
⛓ Sourced · activity log of the department's agents
✎ Action · targeted isolation — approved by your teams
A Blue Lemon Agent departmental multi-agent system coordinates several specialised agents (mail, documents, follow-ups, data entry) under a single point of supervision: consolidated dashboard, activity log, independent control of each agent. It runs with local inference or is hosted in France on a resource isolated by department: the data is never exposed to a foreign service, architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity. Every deliverable remains subject to human approval before it has any effect.
Indicative figures describing our offer — to be confirmed by a pilot on your own scope.
Why coordinate several agents — and why govern how they are supervised
Beyond a single agent, a division gains from having several specialised agents work together. But more agents also means more control points: without overall supervision, visibility is lost.
! The issue
Stacking up independent AI agents, bought separately and poorly integrated, creates a blind spot: who approved what, which agent has access to which data, what happens if one of them goes wrong? Most offers on the market provide no cross-cutting supervision, and host the data outside Europe.
✓ Our answer
A multi-agent system is only worth having if it stays governable: one dashboard, data isolated by scope, traceability for each agent taken individually. Hosting in France, local inference possible, decisions always reserved to your teams: coordination is never paid for with a loss of control.
Governing several agents: supervision & compliance
Coordinating agents must never dilute responsibility. Here is how the architecture protects your data and keeps every agent under control.
Local inference
Each agent can run on a departmental resource: no data leaves the network.
Hosting in France
Otherwise, a dedicated and isolated resource, hosted in France under French law.
Reduced extraterritorial exposure
Architecture designed to reduce exposure to extraterritorial legislation, location alone not being enough to guarantee immunity, even when hosted in Europe, for every agent in the system.
Isolation by agent and by scope
No unintended pooling of data between agents: each scope stays partitioned.
Encryption & controlled access
Encryption in transit and at rest, role-based access (RBAC), strong authentication and logging.
AI Act: governed deployment
Every agent stays strictly in support; no automated decisions; human oversight from end to end.
What depends on the architecture chosen These points are not general guarantees: they are settled deployment by deployment, in the quotation.
- The applicable location is that of the architecture set out in the quotation and verified before commissioning.
- Local execution is announced only for the configuration explicitly described and accepted in the quotation.
- The applicable isolation depends on the deployment mode set out in the quotation; no dedicated isolation is presumed.
- The encryption mechanisms in transit and at rest, their components and key management are those documented for the architecture chosen.
- Roles and permissions are configured and accepted for the identities and systems actually connected.
- The events logged, their content, their retention period and who may access them are defined for the deployment chosen.
See the agent at work
4 real situations, taken from those that come up most often. Pick one: the exchange unfolds as it would in your organisation.
A scripted demonstration. These exchanges show how the agent behaves — its sources, its refusals, what it leaves to your teams. Nothing is sent from this page, no model is queried here, and the matters named are fictional. That is precisely what we promise your data.
The behaviours shown here — monitoring, automation rules, routing and reminders — are configured with you during deployment, from your tools, your rules and your thresholds.
The architecture points named in these exchanges — location, local execution, isolation, encryption, role-based access, logging — are not a guarantee attached to the demonstration: they are those of the architecture set out in your quotation, and verified before commissioning.
· Thirty-one answers contradict each other between two agents. Nobody had seen it: each was answering a different person.
· Fourteen times, one agent cited another agent as a source. That is forbidden here, and I will say why.
· One agent has been producing output nobody consumes for seven weeks.
· Of 2,340 processings, 9 cannot be reconstructed: I cannot say which agent produced what. I flag that rather than rebuild a plausible chain. morning-watch_4-flags.pdf31 contradictions · 14 agent-to-agent citations
⛓ Source · 6 agents, 2,340 processings, chain logs
What I record: 14 times, an agent treated the summary produced by another as if it were the document. In 4 cases, the summary carried an approximation — nothing false, a nuance lost — and the next agent hardened it.
Why this is the risk specific to working in numbers: with one agent, an approximation stays an approximation. With three agents in a chain, it becomes a fact, then a rule, and nobody knows where it came from. None of the three lied.
The rule, and it has no exception: an agent always goes back to the original source. It may read another's output to know where to look — never to cite it.
What it costs: time. Across the 2,340 processings, this rule lengthened 380 of them.
What it prevents: an assertion whose source is an agent. It would be unverifiable, and it would look perfectly sourced. 14-citations_4-hardened.pdfAn agent output is a reading, not a source
⛓ Source · 14 agent-to-agent citations, 4 approximations hardened
What I watch: what each agent receives, what it produces, which source it cited, which version of which reference it applied, and where the output went.
Routing follows who can act: a contradiction between two agents goes to the department lead, with both answers side by side; an agent-to-agent citation to whoever configured the chain; unconsumed output to whoever asked for that agent; a chain that cannot be reconstructed is flagged as such, never rebuilt.
With a weekly chase on contradictions, monthly otherwise. Then a monthly summary: contradictions, agent-to-agent citations, unconsumed outputs, and the share of reconstructible processings.
Three operating rules, and they read as guarantees. Correcting an output is signed by the agent's owner, never by me: a fix applied by supervision makes the discrepancy disappear without its owner ever seeing it, and the agent will do it again, with a supervisor quietly catching it. Suspension is signed too — I prepare it in full: what the agent handled, who received its outputs, what stops with it, and on your mandate I execute it within the minute. And I see only what agents log: a poorly instrumented agent is a blind spot, which I name rather than assume healthy — 9 processings out of 2,340 carry the note "incomplete chain", and it does not come off.
✎ Framework · no correction, no suspension, no rebuilt chain
What I record: 31 cases where two agents answered differently to the same underlying question, put by two different people, days apart. Neither saw the other's answer.
Why it was structurally invisible: each agent is consistent with itself. The contradiction exists only somewhere the two outputs meet — and that place did not exist.
What I found on comparing them: 19 are explained by two references — one applied an earlier version of an internal procedure. 7 by different scopes: both answers were right, about two situations that looked alike. 5 remain real disagreements.
What I do with the 5: I hand them to the department lead with both answers, both sources, and what separates them. I do not say which is right.
What I do with the 19: I flag that two agents are not applying the same version — which is a configuration problem, not a content one, and gets fixed once and for all. 31-contradictions_19-7-5.pdfA contradiction exists only where the two outputs meet
⛓ Source · 31 contradictions, 19 of version, 7 of scope, 5 real
What I can establish: which answers were given, on what date, on what basis. Of the 19 version cases, 11 people received an answer based on a procedure no longer in force.
What stays with the department, and why them: getting back to the eleven. A correction is not a technical notification — it requires knowing what was done with the answer, whether it produced a decision, and what the new one actually changes. None of that is in my logs, and all of it is in the department's files.
What I supply: the list of 11, the answer given, the answer the current version would have given, and the gap between them. Of the 11, 6 gaps changed nothing to the final decision — and saying so avoids eleven letters of which six would have alarmed people for nothing.
What I also flag: how long the outdated version had been applied. Seven weeks. Nobody had any way of seeing it. 11-people_6-with-no-effect.pdfA correction is not a technical notification
⛓ Source · 11 answers on an outdated version, 6 with no effect on the decision
What each processing carries: the agent that produced it, the original source cited — never another agent —, the version of the reference applied, the date, the recipient, and the list of agents traversed before it.
Why the last matters most: a decision passing through three agents looks as if the last one took it. Without the chain, the wrong link gets fixed — and the error comes back the following week by the same route.
The 9 that cannot be reconstructed: two agents produced output without logging their source, following a technical incident. I know they worked, I do not know on what.
What I give on the nine, and what I refuse to write: for eight of them a single source is plausible, and I name it as a hypothesis, alongside the processing, never inside it. The chain itself is not rebuilt: a rebuilt chain looks exactly like a real one, and nobody would know which is which any more.
What I do: those 9 carry the note "incomplete chain", and it does not come off. 2331-reconstructible_9-incomplete.pdfA rebuilt chain looks exactly like a real one
⛓ Source · 2,340 processings, 9 incomplete chains
What I record: one of your six agents has produced a weekly summary for seven weeks. None was opened, none fed another agent, none appears in a handled case.
What that means: it works perfectly. That is exactly the problem — a broken agent announces itself, a useless agent never does and goes on consuming.
What I do not conclude: that it should be stopped. Its output may be read somewhere my logs do not reach, or have been commissioned for a deadline that has not arrived.
What I do: I send the finding to whoever asked for that agent — not to the technical team. The question is not "does it work?" but "what is it for?", and only the requester can answer.
What I measure across the six: the share of outputs actually consumed. It ranges from 96% to zero. It is the most useful figure on this whole dashboard, and the only one nobody ever asks for. 6-agents_from-96-percent-to-zero.pdfA useless agent never announces itself
⛓ Source · 6 agents, share of outputs consumed, 7 weeks
What I can do alone if you open it: suspend an agent whose outputs contradict another's on the same substance, and put it back in service once the contradiction is resolved. What I supply with it: what the agent handled, who received its outputs, and what stops with it.
What stays outside the mandate, and it is not caution: changing an agent's configuration or correcting one of its outputs. A fix applied by supervision makes the discrepancy disappear without its owner ever seeing it — and the agent will do it again, with a supervisor quietly catching it. So I hand the owner the discrepancy, its cause and the correction already drafted: they apply it knowing why, and the agent stops reproducing it.
What it gave: 31 answers contradict each other between two agents. Nobody had seen it, because they never speak to the same person. Each was answering correctly within its own scope.
What supervision cannot do alone: getting back in touch with the people who got the wrong answer. I identify them, I supply what they were told and what they should have been told — the department makes contact, because the department is who answered. suspend-yes_correct-no.pdfWhat opens · why the fix stays with the owner · the 31 contradictions
⛓ Source · 31 contradictions between two agents, people concerned identified
What is kept: the processing chains — agent, source, version, date, recipient, the contradictions found and their outcome, agents' outputs and what becomes of them, the suspensions and their reason, and the non-reconstructible processings.
What "reconstructible" means: I can say, for every output, which agent, which source, which version, on what date, and for whom. Nine processings are not — I list them, because a processing you cannot reconstruct cannot be defended to whoever contests it.
The rule that holds the whole thing up: an agent never cites another agent as a source. An agent's output is a reading, not a source — and a reading that cites another drifts from the original at every step, with nothing signalling it.
What I watch and nobody watches: what the outputs become. One of your six agents has produced a weekly summary for seven months that nobody opens. It works perfectly. what-you-keep_supervision.pdf5 items kept · what outputs become, which nobody watches
⛓ Source · 9 non-reconstructible processings, 1 agent producing into the void for 7 months
Anomaly detection across files: I compare files against each other before a discrepancy reaches management — the same client served at two different rates by two agents, the same document requested twice in two neighbouring files, a stated deadline that contradicts an identical file from the week before. Across 2,340 handlings, anomaly detection raises 68 — 2.9 % — including the 31 contradictions nobody saw, each of the two agents answering a different requester. An anomaly goes to the department, never to the requester, and it carries both files side by side.
Summaries and reporting: every Monday, one page consolidates the activity of the six agents — volume handled, median delay by file type, files waiting and for how long, open anomalies, and the 9 handlings that cannot be reconstructed, which stay in the report until they can be. Building that summary by hand took 4 hours a week; reading it over takes 25 minutes. Over 47 weeks, more than 172 hours given back — more than four 35-hour weeks, rounded down.
Support on the department's human-resources chores, under the responsibility of the department concerned: the table of booked leave and absences to cover, reminders for medical check-ups and authorisations coming up for renewal, the joining file for a new colleague with their access and equipment, answers to internal procedure questions with the originating memo. These are preparations and date reminders. No assessment of a person is produced, and the reason is mechanical: an individual indicator becomes a target, the target distorts what it measures, and the department loses the instrument along with the trust. Whoever signs is named — a signature is not a counter.
The figure that does not flatter me: of my first 120 raised anomalies, 44 were unfounded — 37 %. All of them compared two files from different years, where the rate had changed in between: I was comparing amounts without comparing their dates. Anomaly detection now takes the rate's effective date into account; over the next 300, unfounded ones fell to 7 %, and I re-ran the first 120 to check that no founded anomaly had been lost on the way.
⛓ Sourced · 68 anomalies out of 2,340 handlings, 4 h of summary down to 25 min, 37 % unfounded down to 7 %
Your case is not here? That is exactly what a 15-minute conversation is for. Book the free audit →
Agents that can be coordinated within a single department
Each agent stays independent and is supervised individually; the system brings them into line under one dashboard.
Anomaly detection
Spotting inconsistencies between case files before they reach management.
Summaries & reporting
Periodic consolidation of the various agents' activity for steering purposes.
Human resources
Support on repetitive HR tasks, under the responsibility of the division concerned.
Coordination of the department's sub-agents
Mail, anomalies, reporting, accounting and HR are shared out between dedicated sub-agents, then consolidated under a single point of supervision. Your teams arbitrate and approve every deliverable.
Need to go further?
These agents handle a different business process, with their own owner and their own price. They are added to this one.
Sorting and routing agent
A misrouted request goes round the organisation twice before reaching the right place.
Request sorting and routing agent from 469 € excl. VAT / month Discover the agent →AI executive assistant
A managing director opens the inbox before everyone else and closes it after everyone else.
Executive assistant agent from 652 € excl. VAT / month Discover the agent →Automated reporting
A report that arrives late is no longer any use for deciding.
Automated reporting agent from 499 € excl. VAT / month Discover the agent →AI knowledge base agent
The information exists in your company — but it is scattered across procedures, contracts, an intranet and the memory of a few people.
Document agent (FAQ, knowledge base) from 678 € excl. VAT / month Discover the agent →AI accounting agent
Bookkeeping entry is repetitive, time-consuming and seen as adding no value — yet it has to be exact.
Accounting agent (summaries, anomalies) from 781 € excl. VAT / month Discover the agent →End-to-end HR
Recruitment follows a long path: the need expressed, the vacancy posted, applications examined, interviews arranged, the arrival prepared.
End-to-end HR agent (recruitment → onboarding) from 785 € excl. VAT / month Discover the agent →In 15 minutes we identify the most relevant agent — without oversizing the project.
What coordination changes day to day
Bringing several agents together under shared supervision avoids scattered tools and lost visibility — review time concentrates on the deliverables that matter.
The stages of your AI agent project
Audit & scoping
15 minutes to target the use case with the best return.
Quote or direct sign-up
A catalogue offer is bought online; a specific need gets a costed quote.
Design
We design the agent and its guardrails.
Integration & testing
We connect your tools to the agent, which is itself hosted in France.
Rollout
Going live and training your team.
Operation
Continuous supervision and improvement.
Three packages, one governable system
Three packages, one deployment designed, installed and operated for you. Prices exclude VAT — annual subscription, the time it takes for governance and integrations to settle in for good.
Setup + controlled subscription
- Installation, configuration and training for your teams
- Operation, human oversight, updates and support
- Sovereign hosting in France, a dedicated and isolated resource
All inclusive, no setup fee
- Setup included (installation, configuration, training)
- Operation, human oversight, updates and support
- Sovereign hosting in France, managed end to end
On site, you own it
- Hardware installed on your premises (you own it)
- French / European AI models run locally
- Secure remote maintenance (Pro support included)
Four guarantees that matter for a multi-agent deployment
Your questions, our answers
How many agents can be coordinated in one system?
Can a faulty agent block the others?
Who approves the deliverables the agents produce?
Is the data partitioned between agents?
Does the system fit with our existing tools?
How does deployment work?
Other deployments at company scale
Let's size up the potential for coordination in your department
15 minutes to identify the agents most worth coordinating — hosted in France, supervised, with no commitment.