LAF — autonomous AI development

Several products. One manageable process.

I built Lovtsov Autonomy Framework to develop my own products without spending the day moving instructions between chats. Agents lose context, repeat work, or declare completion before checking the result. Adding more agents does not resolve that on its own.

LAF connects planning, implementation, verification, and decision history. I set the direction, priorities, permissions, and budget. The framework helps prepare a backlog, break down larger work, and execute the accepted queue while preserving progress between sessions.

From a task to a reviewable change

1. Understand and plan. Establish project context, constraints, and acceptance criteria. Decompose larger tasks and their dependencies.
2. Manage the queue. Keep priorities and dependencies in a shared task system, so a stopped session can resume without reconstructing everything.
3. Isolate implementation. Use separate working copies and branches to prevent parallel tasks from silently overwriting one another.
4. Verify. Build, test, review, and inspect the interface. A specific verified revision matters more than an agent’s statement of intent.
5. Hand over. Make changes, evidence, and unresolved decisions available to the owner, who retains authority over release and major product decisions.

The cycle can run in the background within its resource and permission limits. Autonomy is most useful when the work is sufficiently defined and the outcome can be verified.

Beyond the model’s native features

Need	LAF’s approach
Develop several products	Shared queue, separate project contexts, isolated changes
Preserve decisions	Work history, saved state, constraints, reusable practices
See progress	Management dashboard, events, status, Telegram interaction
Control quality	Explicit roles, required checks, failure analysis, verification of the actual code revision
Control resource use	Time and budget limits, provider attribution, stopping when resources are unavailable
Access project knowledge	Tool integrations and a separate RAG assistant

Providers I have used

I have tested LAF with Claude, ChatGPT/Codex, Grok, GLM, and Kimi, combining native capabilities with my own coordination and recovery mechanisms. Connection methods and supported features vary across CLI, API, and provider interfaces. An adoption project includes checking permissions, limits, data requirements, and supported scenarios. The available quota sets the budget for autonomous work.

Working practice and current development

I use LAF on real tasks in my own products and have taken implementations to working versions with it. My AI-assisted and vibe coding practice keeps me responsible for goals, architecture, and acceptance while agents perform a substantial part of the work.

I am improving token efficiency, recovery after interruptions, long-running quality, and onboarding for another operator. For each task, I define the limits of autonomous work and the point at which my acceptance is required.

Work with me

I can demonstrate LAF on a concrete task and discuss adapting it to your projects, provider setup, verification processes, and integrations. I am open to joint development, commercial use, or transfer of the implementation under agreed terms. We define the adoption format around your constraints and needs.

Demonstration: follow a task through delivery

I can show a bounded cycle in an authorised repository: criteria, decomposition, queue selection, an isolated change, checks and a reviewable result. Failure, quota exhaustion and owner decisions are part of the demonstration too.

A pilot starts with one project and a limited task set. Deliverables include a configured workflow, stopping rules and a record of checked changes. Background runtime depends on complexity, integration and quota; leaving it overnight is not a guarantee of any finished product by morning.

Discuss LAF and a demonstration →