liveBook · Manning

Dispatching

As explained in Designing AI Agents

The assignment of a request or unit of work to a processing path or handler based on its properties.

Tool Dispatch routes an action intent to one capability from a controlled candidate set. The runtime first determines which tools are eligible for the role, task stage, object, and granted authority; only then does a model, rule, or router select among them. The runtime owns the registry and the execution path to handlers. section 6.3 section 6.3.1

A dispatch path separates five decisions:

  • Discovery identifies registered capabilities; eligibility narrows them using policy and task state. section 6.3.1
  • Selection chooses an eligible capability or requests clarification; argument validation checks the proposed call against the tool. section 6.3.1
  • Admission checks the exact call against current authority, protected argument provenance, and state freshness before the handler executes. section 6.3 section 6.3.1
  • Acceptance verifies the intended external effect. A registered tool, valid schema, successful return, or green process status alone does not establish success. section 6.3.1 section 6.9
Table 6.5 Discovery, selection, and admission answer different questions
Stage
Question
Owned by
Example rejection
Discovery
What capabilities exist?
Registry or catalog
Tool is not registered
Eligibility
Which capabilities belong to this context?
Policy-aware runtime
Role cannot update production configuration
Selection
Which eligible capability fits the intent?
Model, rule, or router
No candidate fits; ask for clarification
Admission
May this exact call run now?
Execution boundary
Stale base revision or unbound repository path
Acceptance
Did the intended external effect occur?
Task-specific verifier
Wrong deployed version or failed health probe

Use the stage table to distinguish tool discovery and selection from permission to execute and verification of the result. section 6.3.1

Dispatching a deployment

An operations agent may have access to a typed deployment API, an admitted skill, a controlled command, and a legacy GUI. The runtime exposes interfaces appropriate to the task stage, preferring the API or skill and allowing fallback only when justified. Repository identity, base revision, destination, and environment come from protected task state rather than model reconstruction. The selected call still faces admission checks. section 6.3.1

An instruction to use only listed tools is not an executable allowlist. A permitted tool can also carry a wrong identifier or stale state. Fallbacks must be explicit, subject to equivalent authority checks, and recorded; silently switching from an API to a broader shell or browser changes the action boundary. Track eligible-set size, selection precision, admission and stale-state rejections, fallback use, latency, and effect acceptance. section 6.3.4 section 6.3

Binding a template import

For a template import, the host copies exact identifiers from the match receipt into protected import arguments. A dispatcher admits the call against those values before guarded execution; an independent readback checks the resulting group. Admission checks the binding but does not fetch or repair it. Principal authentication and durable recovery remain adapter responsibilities. section 10.7.1 section 10.7.2

Dispatch selects the next capability; it is not a persistent plan. Add planning when dependencies, partial completion, or recovery must survive interruption, and add pre- and postcondition checks when the effect requires them. For a fixed workflow, a direct-binding handler may be sufficient if it enforces the same requirements and passes the same gates. section 6.3.4 section 10.7.1 section 6.6

How Designing AI Agents covers it

How deep it goes
Details eligibility, selection, validation, admission, execution, and acceptance, with failure modes, measures, and composition guidance.
Examples
Uses deployment interface selection and a receipt-bound template import to illustrate eligibility, protected arguments, and effect verification.
What you can do after
Apply dispatching to constrain tool selection, validate proposed calls, and assess their external effects.
How it connects
Connects dispatching to capability systems, API gateways, planning, prompt chaining, and guardrails.

Related concepts

This explainer was generated by Manning's knowledge graph from the text of Designing AI Agents. The section links point to the parts of the book it draws on.