Skip to content

GCRI · Intelligenza artificiale e sistemi autonomi

Sistemi agentici

Agentic AI sistemi 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.

01Registrare

Agentic systems on the record.

02La domanda

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.

Meccanismo

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.

Meccanismo

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.

Meccanismo

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.

Meccanismo

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.

03Monitoraggio

Agent oversight console.

La schermata che un gruppo utilizzerebbe per seguire questo tema: misure, stati e decisioni. I nomi sono esempi e i valori sono illustrativi.

Agent oversight console

Agentic systems · Agent sandbox · Session AG-6627

Human approval pendingGCRI · Intelligenza artificiale e sistemi autonomi

Tool calls

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

Politiche

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

Actions by approval class

050100count this weekAutomaticoRegistratoApprovatoDeniedEscalation effettuata
Agent oversight console: the screen’s structure, terms and states for Agentic systems.Nomi di esempio e valori illustrativi; un progetto, non un registro né un sistema attivo.
04Tra i domini

Agents that act need guardrails in every domain they touch.

Agentic system studies are designed for the serie Exponential Technology. 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.
05Metodi ed evidenze

How agentic systems is modelled, measured and tested.

Il modello, le sue ipotesi, i suoi dati e la sua incertezza, nella forma in cui uno scienziato o un ingegnere li verificherebbe. Le strutture seguono metodi pubblicati. I valori rappresentati sono illustrativi, non risultati.

Modello di sistema

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

++−+Intelligenza artificialeDatiComunitàEconomiaR di rinforzoB bilanciamento
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.
Principi guida FAIR · GO FAIR e FORCE11W3C PROV · World Wide Web ConsortiumGuidance on consistent treatment of uncertainties · IPCCCEOS Analysis Ready Data · Comitato sui satelliti per l'osservazione della Terra
06Indicatori

Cosa viene misurato e con quale accuratezza.

Ogni indicatore riporta la sua unità, la sua cadenza, il tipo di fonte su cui si basa e l'etichetta di evidenza che un valore riporterebbe. Non sono mostrati valori: un programma li fornisce sotto la propria autorità.

IndicatoreUnitàCadenzaTipo di fonteEtichetta
Task success rate on agent benchmarkspercent of tasks completed correctlyper evaluationAgent benchmark resultsRIPORTATO
Actions requiring human approvalpercent of consequential actionsper deploymentDeployment configurationDICHIARATO
Prompt injection success ratepercent of attempts succeedingper evaluationSecurity testingOSSERVATO
Agent incidentscount by severitymensileIncident reportsRIPORTATO
Actions logged and attributablepercent of agent actionsper deploymentAudit logsOSSERVATO
Mean task length before human interventionsteps or minutesper evaluationEvaluation logsOSSERVATO
07Fonti

Da dove provengono le evidenze.

Set di dati pubblici e registri istituzionali specifici, ciascuno con la sua classe di licenza. Le fonti soggette ad accordo sono utilizzate solo alle relative condizioni e mai ripubblicate.

FonteCosa fornisceClasse di licenza
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 sistemiOpen, licence to check
Developer system cards describing agent safeguardsDeveloper-reported agent evaluations and controlsPublic documents, terms per developer
08Decisioni

Chi decide, e a cosa servono le evidenze.

Studio Intelligenza artificiale e sistemi prepares evidence. The decisions below are taken by the authorities named, under their own rules, and never by the platform.

Chi decideLa domandaEvidenze fornite da questa pagina
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
Autorità di regolamentazioneWhat oversight to require for autonomous AI actionsOversight design evidence

Output prodotti da questo argomento

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

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?
10Meccanismi

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. sistemi that coordinate several agents add further failure modes, as agents pass errors to each other or interact in unforeseen ways.

11Metodi e dati

Designing for safe execution.

  1. 01

    Least privilege

    Granting only the permissions a task needs, with sandboxes and spending or action limits, drawing on cybersicurezza e crittografia.

  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 Ingegneria della valutazione e dell'assicurazione dell'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.

Gli organismi e i documenti alla luce dei quali si legge questa pagina. Stato ed edizioni cambiano; ogni link porta all'organismo emittente.

14Domande

Domande e risposte.

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.
RapportiContribuisci

Contribuisci con evidenze da questa pagina.

Rapporti 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.

GCRIIntelligenza artificiale e sistemi autonomiTecnologia esponenzialeEvidenze di Labs
Intelligenza artificiale e sistemi autonomi

Porta un metodo, un set di dati o una domanda.

Informazioni e richieste sono aperte oggi. Una richiesta riceve un riferimento e indica il passo successivo; candidature, pagamenti, assegnazioni di calcolo e registrazioni si aprono solo quando esistono il team, le condizioni e la capacità di revisione.