Skip to content

GCRI · Intelligence artificielle et systèmes autonomes

Systèmes agentiques

Agentic AI systèmes pursue goals by planning and taking actions: calling tools, browsing, writing and running code, or operating other software, sometimes over long tasks with little supervision. Autonomy brings usefulness and new failure modes, from compounding errors to misuse of granted permissions. This page sets out how agents fail, how their actions can be bounded and overseen, and what evidence serves those who build, deploy and govern them.

01Registre

Agentic systems on the record.

02La question

Agentic systems: how the risk develops.

Tool access, delegated authority and safe execution. The dossier below sets out how the risk develops, what is measured, where the data come from, how it is modelled and who decides.

Mécanisme

Errors compound over steps

Each step in a long task can go wrong, and an early mistake shapes later actions. Small per-step error rates add up over many steps.

Mécanisme

Permissions define the blast radius

An agent can only do harm with the access it has been given. Broad credentials, write access and payment rights turn errors or manipulation into real-world consequences.

Mécanisme

Untrusted inputs steer behaviour

Agents read web pages, documents and messages that may contain instructions. Prompt injection can redirect an agent to act against its user.

Mécanisme

Oversight must be designed in

Humans cannot review every action of a fast agent. Approval steps for consequential actions, logging and the ability to stop and roll back are design choices.

03Suivi

Agent oversight console.

L'écran qu'une équipe utiliserait pour suivre ce sujet : ses mesures, ses états et ses décisions. Les noms sont des exemples et les valeurs sont illustratives.

Agent oversight console

Agentic systems · Agent sandbox · Session AG-6627

Human approval pendingGCRI · Intelligence artificielle et systèmes autonomes

Tool calls

  1. 10:34Infosearch(documents) · read-only
  2. 10:51Inforun_model(flood_nowcast) · sandbox
  3. 10:51Jalonsend_message(external) · needs approval
  4. 10:53Errorwrite_file(outside workspace) · denied

Politiques

Périmètre
workspace only
External actions
human approval
Spend limit
43 units
Kill switch
Armed

Actions by approval class

050100count this weekAutoConsignéApprouvéDeniedRemonté
Agent oversight console: the screen’s structure, terms and states for Agentic systems.Noms donnés à titre d'exemple et valeurs illustratives ; une conception, non un enregistrement ni un système en service.
04À travers les domaines

Agents that act need guardrails in every domain they touch.

Agentic system studies are designed for the série Technologies exponentielles. The OWASP Top 10 for LLM applications lists failure modes such as excessive agency and prompt injection; the page describes them without exploit detail.

  • Incident reports involving agents.
  • Designs for oversight and permission scoping.
05Méthodes et preuves

How agentic systems is modelled, measured and tested.

Le modèle, ses hypothèses, ses données et son incertitude, sous la forme qu'un scientifique ou un ingénieur vérifierait. Les structures suivent des méthodes publiées. Les valeurs représentées sont illustratives, ce ne sont pas des résultats.

Modèle de système

The couplings between the systems in agentic systems: signs of influence, a reinforcing and a balancing loop, and cross-links where stress jumps.

++−+Intelligence artificielleDonnéesCommunautésÉconomieR renforçanteB équilibrage
Causal loop diagram from the page’s systems; signs show the direction of influence. Illustrative structure for a system dynamics or network model, not estimated couplings.
Principes directeurs FAIR · GO FAIR et FORCE11W3C PROV · World Wide Web ConsortiumOrientations sur le traitement cohérent des incertitudes · GIECCEOS Analysis Ready Data · Comité sur les satellites d'observation de la Terre
06Indicateurs

Ce qui est mesuré, et avec quelle qualité.

Chaque indicateur porte son unité, sa périodicité, le type de source qui le sous-tend et l'étiquette de preuve que porterait une valeur. Aucune valeur n'est affichée : un programme les fournit sous sa propre autorité.

IndicateurUnitéFréquenceType de sourceÉtiquette
Task success rate on agent benchmarkspercent of tasks completed correctlyper evaluationAgent benchmark resultsDÉCLARÉ
Actions requiring human approvalpercent of consequential actionsper deploymentDeployment configurationDÉCLARÉ
Prompt injection success ratepercent of attempts succeedingper evaluationSecurity testingOBSERVÉ
Agent incidentscount by severitymensuelIncident reportsDÉCLARÉ
Actions logged and attributablepercent of agent actionsper deploymentAudit logsOBSERVÉ
Mean task length before human interventionsteps or minutesper evaluationEvaluation logsOBSERVÉ
07Sources

D'où proviennent les données probantes.

Jeux de données publics et registres institutionnels nommés, chacun avec sa catégorie de licence. Les sources fournies dans le cadre d'accords ne sont utilisées que selon leurs conditions et ne sont jamais republiées.

SourcesCe qu'elle apporteCatégorie de licence
OWASP Top 10 for LLM Applications and agentic security guidanceRisks such as prompt injection and excessive agencyOpen, CC BY-SA 4.0
Published agent benchmarks such as SWE-bench and related suitesTask-based evaluation of agentsOpen, licence per benchmark (to check)
NIST AI Risk Management FrameworkPractices for governing and measuring AI riskPublic domain, US government
AI Incident DatabaseReported incidents involving AI systèmesOpen, licence to check
Developer system cards describing agent safeguardsDeveloper-reported agent evaluations and controlsPublic documents, terms per developer
08Décisions

Qui décide, et à quoi servent les données probantes.

Studio Intelligence artificielle et systèmes prepares evidence. The decisions below are taken by the authorities named, under their own rules, and never by the platform.

Qui décideLa questionDonnées probantes servies par cette page
Deploying organisationsWhich actions an agent may take alone and which need approvalPermission patterns; evaluation results
Security teamsHow to protect agents from manipulationInjection testing; incident evidence
RégulateursWhat oversight to require for autonomous AI actionsOversight design evidence

Résultats que produit ce thème

  • Permission and approval design patterns for agents
  • Agent evaluation protocols covering long tasks
  • Prompt injection test suites for deployments
  • Logging and rollback requirements
09Questions essentielles

What designers and evaluators ask.

  • What actions can the agent take, with whose authority, and which are reversible?
  • How does it behave when instructions are ambiguous, conflicting or malicious?
  • Can a human see what it is doing, understand why and stop it in time?
  • How do errors compound over long sequences of steps?
10Mécanismes

How agents fail.

An agent combines a model’s reasoning with permissions in the world. Small errors in early steps can compound across a long task. Agents can misinterpret goals, pursue them in unintended ways or take shortcuts that satisfy a metric but not the intent. Content the agent reads, such as a web page or document, can contain instructions designed to hijack it, an attack known as prompt injection.

The more permissions an agent holds, the larger the potential harm. systèmes that coordinate several agents add further failure modes, as agents pass errors to each other or interact in unforeseen ways.

11Méthodes et données

Designing for safe execution.

  1. 01

    Least privilege

    Granting only the permissions a task needs, with sandboxes and spending or action limits, drawing on cybersécurité et cryptographie.

  2. 02

    Human oversight points

    Requiring confirmation for consequential or irreversible actions. Too many prompts lead to rubber-stamping, so placement matters.

  3. 03

    Task-based evaluation

    Testing agents on realistic multi-step tasks in controlled environments, using Ingénierie de l'évaluation et de l'assurance de l'IA. Simulated environments miss some real-world complexity.

  4. 04

    Logging and audit

    Recording actions and reasoning traces so incidents can be reconstructed. Traces may not fully reflect how a model reached a decision.

13Where this page’s references

Where this page’s references lead.

Les organismes et documents à l'aune desquels cette page se lit. Les statuts et éditions changent ; chaque lien mène à l'organisme émetteur.

14Questions

Questions et réponses.

What is an AI agent?
An AI system that pursues goals by planning and taking actions, such as using tools, browsing or running code, with some degree of autonomy.
What is prompt injection?
An attack in which instructions hidden in content the AI reads, such as a web page or document, cause it to act against its user’s intent.
How can agent risk be reduced?
By limiting permissions to what the task needs, requiring approval for consequential actions, logging all actions, testing against manipulation and keeping the ability to stop and undo.
Should agents ever act without human approval?
For low-consequence, reversible actions within tight limits, often yes. For consequential or irreversible actions, the design question is where and how a human confirms.
Can agents be evaluated before deployment?
Partly, in controlled environments with realistic tasks. Monitoring after deployment remains necessary because real environments are more varied.
Does the platform build agents?
Its work is evaluation and method. Any build work for Consortium projects runs through Foundry under its own governance.
RapportsContribuer

Apportez des éléments de preuve depuis cette page.

Rapports is the Consortium’s quarterly publication. What you know about this subject can become evidence, a case or a correction in the next edition, credited to you and published under CC BY 4.0 where it is your own work.

GCRIIntelligence artificielle et systèmes autonomesTechnologies exponentiellesPreuves des Labs
Intelligence artificielle et systèmes autonomes

Apportez une méthode, un jeu de données ou une question.

L'information et les demandes sont ouvertes dès aujourd'hui. Une demande reçoit une référence et indique l'étape suivante ; candidatures, paiements, allocations de calcul et inscriptions n'ouvrent que lorsque l'équipe, les conditions et la capacité d'examen existent.