AI & agents

Built-in agent actions

Gives an agent a fixed set of typed, bounded tools instead of raw engine or shell access.

BetaEngine edition 5f8ea3af

What it does

The runtime registers 18 actions: 13 built-in observation and memory tools, 4 cloud-provider actions and one delegation action. Each declares a JSON schema and an effect, and the approval policy is derived from the effect. GET /admin/actions lists every action with its schema, effect and whether approval is required.

How it works

The thirteen built-ins are stats, list_collections, count_records, list_records, list_files, list_deployments, list_db_sinks, list_kinds, log_tail, health_check, recall_similar_runs, rag_over_log and recall_memory. They are bounded by schema and by code: list_records returns at most 100 records, log_tail at most 200 one-line summaries of 200 characters, rag_over_log at most 20 past runs with 400-character excerpts, and recall_memory clamps top_k and graph hops the way the memory route does. None of them writes a record.

rag_over_log and recall_similar_runs run a BM25 search over earlier agent runs, so an agent can ground an answer in prior work. recall_memory calls the fused episodic and semantic recall of the agent memory capability.

The four provider actions are list_provider_instances, deploy_to_provider, destroy_provider_instance and create_cloudflare_hyperdrive. They use the provider credential vault and cloud-provider code; the first reads without changing anything, and the other three declare an external or destructive effect, which makes them approval-gated. delegate_to_agent runs another named agent, with a depth limit of three.

An action declares its effect once. An action with no side effect, and a reversible write, needs no approval; destructive-write and external-effect require it. An agent can call just the actions on its allow-list, and an empty list grants nothing.

Where you reach it

  • HTTPGET /admin/actions (name, description, input_schema, effect, requires_approval, risk)

Limits

  • Treat agents as an operator facility and grant each one just the actions it needs.
  • The stats action reports the data log path to the model, and therefore into the stored run.
  • The description of rag_over_log says semantic search; it is BM25 keyword search over run text.
  • Provider actions depend on credentials registered in the provider vault.
  • There is no action that writes records or files. Writes happen through the approval-gated provider actions and nowhere else.

Evidence

  • On this siteAI features catalogue
  • TestsA test checks that the registry registers every built-in; 3 provider-action tests check approval gating and denial before execution; 3 tests cover the memory-recall action

Test coverage is listed as present in the engine repository at this edition. This page does not report test results.

KynetraDB is in private preview

Request a preview to talk through the limits above against your own workload.