The platform

Keep

An experimental platform for doing useful work with AI agents—and examining the conditions under which that work can be trusted.

Active development; public preview in preparation

01 / A shared foundation

Keep has its own native execution machinery. Its design spans repository work, persistent knowledge, reusable procedures, delegated authority, and evidence about outcomes. It is intended for both individual operators and more complex organizational environments, without requiring a particular external coding-agent product.

Support for different providers and integrations is qualified separately. An available interface is not the same as a verified deployment, and not every planned capability is complete.

What the platform brings together

These areas are at different stages of implementation and qualification.

Native execution and acceptance: Repository investigation, proposed changes, testing, and separation between producing a result and authorizing its acceptance.
Persistent knowledge: Retained sources, scoped access, corrections, and checks on whether previously supplied information remains eligible for use.
Skills and adaptation: Mechanisms for proposing, evaluating, promoting, and retiring reusable procedures or other changes to agent behavior.
Coordination and recovery: Controlled admission of operations, shared-resource accounting, and explicit treatment of effects whose outcomes remain uncertain.
Delegation and privacy: Limited grants of authority and configured controls on what information reaches particular execution or provider boundaries.
Evaluation and audit evidence: Outcome checks and inspectable records, with clear limits on what they establish.

02 / Development evidence

Recorded development workflows

Historical development evidence, with a defined scope for each result.

Governed repository work

Recorded installed workflows have exercised proposal, approval or refusal, merge, restart, and revert. The purpose is to keep producing a patch distinct from deciding whether it should be accepted.

Scope: defined development workflows, not proof of correctness for arbitrary repositories or production deployments.

Coding with corrected information

An installed temporal-memory task used retained information, distinguished an earlier value from a later replacement, and produced a repository repair checked against separately defined assertions.

Scope: an authored task family, not a broad continual-learning result or evidence of the model's internal reasoning.

Recovery when an outcome is uncertain

Controlled experiments use a durable local service to examine what happens when an operation may have succeeded but its acknowledgement is lost. The tested corrections retain the relevant obligations rather than treating an error as proof that nothing happened.

Scope: synthetic, named execution paths and tested restart boundaries—not real payment-provider qualification or a universal recovery guarantee.

These summaries refer to recorded development experiments. Versioned public artifacts and reproduction instructions will be linked after release review. They do not constitute independent replication or a claim of superiority over other systems.

03 / In development

Current release status

Keep is under active development. A reviewed public research preview is in preparation. The preview's supported configurations, demonstrations, and limitations will be documented with the release.

Keep is not being presented here as a fully qualified production system. General learning gains, broad interoperability, and comparative advantages require further evaluation.

Explore the research questions