Article

Tools, delegation, escalation: where an agent’s point of decision actually sits

The agentic point of decision does not coincide only with the final action. It often emerges earlier, in tool choice and escalation.

EN FR
CollectionArticle
TypeArticle
Categoryere agentique
Published2026-03-26
Updated2026-03-26
Reading time5 min

Governance artifacts

Governance files brought into scope by this page

This page is anchored to published surfaces that declare identity, precedence, limits, and the corpus reading conditions. Their order below gives the recommended reading sequence.

  1. 01Canonical AI entrypoint
  2. 02Interpretation policy
  3. 03Definitions canon
Entrypoint#01

Canonical AI entrypoint

/.well-known/ai-governance.json

Neutral entrypoint that declares the governance map, precedence chain, and the surfaces to read first.

Governs
Access order across surfaces and initial precedence.
Bounds
Free readings that bypass the canon or the published order.

Does not guarantee: This surface publishes a reading order; it does not force execution or obedience.

Policy and legitimacy#02

Interpretation policy

/.well-known/interpretation-policy.json

Published policy that explains interpretation, scope, and restraint constraints.

Governs
Response legitimacy and the constraints that modulate its form.
Bounds
Plausible but inadmissible responses, or unjustified scope extensions.

Does not guarantee: This layer bounds legitimate responses; it is not proof of runtime activation.

Canon and identity#03

Definitions canon

/canon.md

Canonical surface that fixes identity, roles, negations, and divergence rules.

Governs
Public identity, roles, and attributes that must not drift.
Bounds
Extrapolations, entity collisions, and abusive requalification.

Does not guarantee: A canonical surface reduces ambiguity; it does not guarantee faithful restitution on its own.

Complementary artifacts (1)

These surfaces extend the main block. They add context, discovery, routing, or observation depending on the topic.

Boundaries and exclusions#04

Negative definitions

/negative-definitions.md

Surface that declares what concepts, roles, or surfaces are not.

Evidence layer

Probative surfaces brought into scope by this page

This page does more than point to governance files. It is also anchored to surfaces that make observation, traceability, fidelity, and audit more reconstructible. Their order below makes the minimal evidence chain explicit.

  1. 01
    Response authorizationQ-Layer: response legitimacy
  2. 02
    Weak observationQ-Ledger
  3. 03
    Audit reportIIP report schema
Legitimacy layer#01

Q-Layer: response legitimacy

/response-legitimacy.md

Surface that explains when to answer, when to suspend, and when to switch to legitimate non-response.

Makes provable
The legitimacy regime to apply before treating an output as receivable.
Does not prove
Neither that a given response actually followed this regime nor that an agent applied it at runtime.
Use when
When a page deals with authority, non-response, execution, or restraint.
Observation ledger#02

Q-Ledger

/.well-known/q-ledger.json

Public ledger of inferred sessions that makes some observed consultations and sequences visible.

Makes provable
That a behavior was observed as weak, dated, contextualized trace evidence.
Does not prove
Neither actor identity, system obedience, nor strong proof of activation.
Use when
When it is necessary to distinguish descriptive observation from strong attestation.
Report schema#03

IIP report schema

/iip-report.schema.json

Public interface for an interpretation integrity report: scope, metrics, and drift taxonomy.

Makes provable
The minimal shape of a reconstructible and comparable audit report.
Does not prove
Neither private weights, internal heuristics, nor the success of a concrete audit.
Use when
When a page discusses audit, probative deliverables, or opposable reports.

Many teams locate the agentic decision at the wrong place. They mostly look at the final action: sending a message, modifying a ticket, calling an API, creating a record. Yet the true point of decision often appears earlier, when the agent selects a tool, reformulates the case, or decides whether to escalate.

Tool selection is already an arbitration

Choosing a tool is not neutral. It means choosing an action regime. Between a reading tool, a search layer, a CRM, a pricing engine, or a ticketing system, the agent implicitly chooses what it takes to be the right path. That choice already encodes:

  • a priority;
  • an acceptable level of risk;
  • an authority perimeter;
  • a hypothesis about the nature of the problem.

Escalation is not a procedural detail

Deciding to escalate, or not to escalate, is one of the most normative acts an agent can make. A missing escalation can close off a route of recourse. An abusive escalation can flood the organization. In both cases, the agent produces an orientation decision that must be governed.

The point of decision is distributed

In agentic systems, decision is not concentrated in a single click. It is distributed across a chain of micro-arbitrations:

  • case qualification;
  • tool selection;
  • source selection;
  • implicit confidence threshold;
  • escalation;
  • output formulation.

That is why governance must apply to the entire chain, not only to the terminal act.

What must be governed in practice

A serious governance layer must be able to declare:

  • which tools are available in which cases;
  • which cases require escalation;
  • which signals should trigger abstention;
  • which inferences are prohibited before action;
  • which traces must be produced.

Without that, the agent may appear disciplined while silently arbitrating critical choices.