Knowledge about externally owned tools

Responsibility boundary#

Tools belong to the application and the company whose tasks they serve. The application owns their code, packages, dependencies, maintenance, release selection, business permissions and operational lifecycle. Its infrastructure runs tools with appropriate isolation. A lightweight agent client need not host a compiler, container runtime or execution server.

QilbeeDB stores knowledge about tools: purpose, applicability, instructions, preconditions, limitations, examples, observed outcomes and provenance. External version references associate knowledge with the implementation that produced an observation. They do not transfer ownership of that implementation to the bank.

Responsibility Owner
Create, repair, package and publish executable tools Agent application and company
Select a tool for a business task and authorize its execution Agent application under company policy
Dispatch, cancel, recover and account for execution Application execution infrastructure
Persist and retrieve scoped knowledge and its source history QilbeeDB
Evaluate recorded knowledge under an explicit evidence policy QilbeeDB, with observations and evaluation context supplied externally
Determine whether an external version is available and permitted now Agent application and execution infrastructure

Persisting evidence does not make the database the owner of the business process. A knowledge qualification decision is not permission to publish or execute code.

Existing capabilities and compatibility#

The Learned Tool API currently exposes source-text artifacts, executor profiles and development outcome records. These existing interfaces remain documented for compatibility. They are not prerequisites for ordinary memory or procedural knowledge, and their existence does not establish the intended ownership of future capabilities. They do not dispatch programs or publish executable releases.

The previous architecture that assigned development, packaging and an execution gateway to QilbeeDB is superseded. Any migration or removal of existing interfaces requires usage assessment, an explicit compatibility plan and agreement with affected consumers. This document does not remove endpoints or migrate records.

The Procedural Learning API already supports instructions, source references and immutable evaluation contexts. Contexts can identify external tool revisions without storing executable code. An exact revision string is a declared identity, not independent proof of an external artifact's existence, integrity or current availability. A schema or catalog revision identifies an offered interface; it must not be presented as an executable implementation revision.

Evidence and current observations#

Useful knowledge should identify its originating observations and applicable context. Comparisons must distinguish the baseline, candidate instructions, model, tools, dataset, harness, environment and permissions. A standalone tool comparison and an entire agent-task comparison answer different questions; do not silently substitute one identity or claim for the other.

A historical receipt records what was accepted at that time. A current read must distinguish qualification state from the current validity of source memories. Updated, rejected, deleted or expired evidence can invalidate reuse. Expiry may occur without a change-feed event. Incomplete traversal, missing evidence and unknown resource consumption must remain explicit, never counted as success.

The additional generic contract binding procedural knowledge to exact memory source revisions is under cross-team review. Its fields and endpoints are not available contracts. The unpublished artifact-dependent candidate design is suspended; no code registration or executor profile should be required by its replacement.

Knowledge inspection does not reserve execution or eliminate the race between inspection and use. Current business authorization and execution recovery remain application responsibilities. The database must not fetch arbitrary external locations or invoke code merely to inspect a knowledge reference.

Ownership review before implementation#

Before adding a capability, identify the user need, existing owner, alternatives, data and decision authority, security boundaries and lifecycle responsibilities. Database persistence alone is not sufficient justification for database ownership.

When the change affects a shared boundary, both teams must explicitly agree on ownership, the integration contract, acceptance evidence and compatibility impact before implementing the dependent change. Silence is not agreement. Independent investigation and internal work may continue. Reopen the review when scope changes.

Validate source integrity, tenant and private-subject isolation, current state, recovery and bounded work independently of retrieval quality. Claims of better tool selection or agent capability additionally require comparative task evidence.