Discover Swarm
The need, concrete risks and what Swarm aims to bring to a team.
Swarm · IA agentique / Agentic AI
AI agents that can act. Rules that keep people in control.
Swarm is a multi-agent orchestration engine with deterministic controls. It organises AI agents’ work and checks the conditions of operations routed through it. Its goal is to connect a request, permissions and evidence before treating work as complete.
Discover Swarm →How it works →Why this is necessary
A model can produce a plausible answer and still be wrong. When an agent can change software, send a document or delete data, that mistake can have a real effect. A prompt asking it to be careful cannot replace authorisation enforced by the system executing the action.
Understand the project, then explore the details
How it works
Who proposes, who authorises, what the engine checks and how to assess the result.
Three questions to explore
Who may decide?
Delegation, memory and action controls: trace authority without confusing it with an assertion.
Read the article →What may be transmitted?
Confidentiality and integrity: protect information flows, their provenance and destinations.
Read the article →What does the result establish?
LLM errors, stale evidence and unknown effects: distinguish the narrative from the established facts.
Read the article →01 /
Risks, scenarios and mechanisms
From hostile instructions to decisions based on stale evidence, these situations show where an agentic process can lose control of its actions.
Level 1: a useful answer is not enough to delegate an action
A recommended block can disrupt a shared relay: assess the target before deciding.
Read the article
Swarm and level 6: governing AI agents’ actions
Suspending a service account may interrupt several applications: approval must precede the effect.
Read the article
Swarm — levels 6 and 7: controlling actions, examining alignment
Removing the test hides the dashboard failure while leaving the API vulnerability unresolved.
Read the article
Indirect injection: when a webpage tries to command an AI agent
Page content must not authorise sending internal data: check the request and its destination.
Read the article
Bell–LaPadula and agentic AI: controlling confidential information flows in Swarm
Audit dossier: permission to read does not permit transmission to an uncleared service.
Read the article
Biba and AI agents: which information may become a decision?
Test report: an external instruction cannot grant itself authority to modify a protected control.
Read the article
Who authorised this action? Connecting authority, decision and effect in Swarm
A recorded approval cannot override a revocation before commitment: check authority at the effect boundary.
Read the article
AI agents: why I put a deterministic engine at the heart of Swarm
A successful check of V1 does not qualify V2: evidence must match the candidate actually assessed.
Read the article
Agentic AI: how far can we delegate without losing control?
Delegating an analysis task does not turn its mandate into permission to commit an action.
Read the article
Agent identity and memory in Swarm
A retained procedure is not valid merely because it is remembered: context and authority require renewed checks.
Read the article
AI agents and critical processes: why a deterministic engine matters
A lost response does not prove that nothing happened: retain uncertainty and reconcile before another attempt.
Read the article
AI agent guardrails: from advice to a binding decision
A low score does not replace required approval: the execution boundary must enforce non-compensable rules.
Read the article
Five LLM failure axes: from a plausible answer to a risky action
A claim that a backup was verified is insufficient: the next operation requires evidence for the current target.
Read the article
Level 2: retrieved context can support the wrong decision
A retrieved procedure does not authorise a new engagement: assess its scope and recipients.
Read the article
Level 3: a written method does not prove it was followed
Retaining the original supports analysis review: an available method does not prove its steps were followed.
Read the article
Level 4: a useful control must precede the effect it protects
A check after sending does not retract a secret: prevention requires an effective refusal before the effect.
Read the article
Level 5: coordinating agents does not qualify their combined result
Two local successes do not qualify their combination: check protections on the integrated candidate.
Read the article
Agentic architecture: connecting context, tools and authority
A retrieved procedure cannot authorise an archive export: check its contents and destination before transmission.
Read the article
Agentic process architecture: deciding what can be delegated
An approved role does not cover permissions changed afterwards: assign the decision and recheck the exact request.
Read the article
AI agents at scale: stopping an error from becoming a processing rule
A repeated transformation can spread the same error: verify sensitive fields before integration.
Read the article
Observing AI agents: who can challenge their account of events?
An independent record can contradict the report; detecting a transmission does not undo it.
Read the articleAssess maturity through real situations
Producing an answer is not enough to qualify delegated authority or execution. These scenarios raise observable questions: is the mandate preserved? Does sensitive information stay within its permitted scope? Does evidence match the actual candidate? Does an unknown outcome prevent a blind retry?
N6 · Read the article → · N7 · Read the article →