Autonomous systems
DevAuto / active research
DevAuto
AI that can keep working because the project—not the chat—is the durable unit of context.
Unit
Project
Memory
Durable
Tools
Controlled
Clients
Many
System boundary
A runtime, not a prompt wrapper.
DevAuto is being designed around projects with their own files, memory, agent configuration and tool permissions. Interfaces can change; the project state remains.
Project container
The filesystem, memory and work state belong to a project rather than a transient conversation.
Agent runtime
Planning and execution are separated from provider-specific chat details.
Controlled tools
Tools expose explicit capabilities and boundaries instead of giving an agent unlimited ambient access.
Durable memory
Useful state survives across sessions so long tasks do not reset when context windows do.
One project / many surfaces
The intelligence should outlive the interface.
CLI
Fast operational control
Desktop
Long-running project workspace
VS Code
Work next to the codebase
Durable work loop
Short context. Long ambition.
Goal
A bounded project objective.
Plan
Work is decomposed into explicit tasks.
Act
Tools operate inside allowed capabilities.
Verify
Outputs are checked against immediate evidence.
Persist
State is written back for the next iteration.
Research questions
Where autonomy actually breaks.
How should an agent know when to ask, stop, retry or hand control back to a human?
What project memory is useful enough to persist without becoming noisy or unsafe?
How can tool contracts remain stable while models and clients change?
Can immediate validation make bounded autonomous work reliable with much smaller context windows?