The UAIS approach
Giving agents more work without losing control.
UAIS calls this Agentic Software Engineering: an engineering category where AI agents are brought into the production process through contracts, approval, and independent verification. As agent authority grows, what reduces risk is not more human approval — it is a better-defined contract, bounded authority and independent verification.
A contract is a working surface, not documentation.
Every task moves through the same flow: request, scope, acceptance criteria, authority, output and evidence. If one of these steps is missing, the task does not proceed.
In UAIS, a contract is not just documentation. It is the working surface that defines what an agent can read, what it can change, what evidence it must produce, and where it must hand back to a human.
Design principle
HumanGate: approval stays with the human.
Not every action carries the same weight. UAIS classifies actions by risk class; an easily reversible read and a hard-to-reverse delete or publish do not run at the same level of automation.
Actions that require approval are defined in advance. Granted authority is not general or permanent — it is scoped to a specific task and context, and is often single-use.
The producer and the reviewer are not the same session.
The agent session that produces a decision or code is separated from the one that verifies it; a producer does not approve its own output on its own judgment.
Verification is not a subjective “looks fine” judgment: test oracles and deterministic checks run and produce a pass or fail result.
Every decision is written to a permanent ledger.
The request, the decision taken, the reason for rejection and the approval granted are recorded with a timestamp; this record is not altered afterward.
Version and cost history are read from the same ledger: which change ran when, and under what budget, can be reconstructed later.
Institutional memory does not disappear with the session.
Context does not reset when an agent session ends. A new task builds on prior decisions and contracts.
Contract documents
CLAUDE.md and project contracts are the first context an agent reads the moment it enters a repository.
RAG and knowledge graph
A hybrid retrieval layer combines vector search, a knowledge graph and keyword search to surface the relevant context.
Atom registry
Verified, versioned units of knowledge are sealed; knowledge validated once does not need to be rediscovered.
Decision records
Prior approvals and rejections are read back as context when a new task starts.
Principles and products
The principles below are not embodied in a single product — they show up across multiple products in the UAIS family.