ZLoop

ZLoop is a vendor-neutral, production-oriented framework for building AI agent systems that improve their own work through bounded feedback loops.

Core lifecycle

DISCOVER → PLAN → EXECUTE → VERIFY → REVIEW → REPAIR → VERIFY → SHIP

Every run is constrained by explicit budgets, stop conditions, permission boundaries, independent verification, persistent memory, and human handoff.

Principles

• Closed loops first.
• Maker/checker separation.
• No unbounded retries.
• Persist state outside model context.
• Mutations require authorization and idempotency.
• Verification decides completion, not the executor.
• Cost/context are first-class resources.
• Parallel agents use isolated worktrees.
• Every material action is auditable.

Components

• Orchestrator, Discoverer, Planner, Executor, Verifier, Reviewer, Repairer
• Memory Manager and Cost/Context Governor
• reusable skills under skills/
• JSON contracts under schemas/
• policy-as-code documentation under policies/
• provider-neutral Python reference runtime under src/
• coding, research, content and fleet examples

Quick start

python -m unittest discover -s src -p 'test_*.py' -v
python src/loop_engine.py

Replace the demo adapter with adapters for your preferred LLM/model gateway, GitHub/CI, issue tracker, database, staging API, or other connector.

Acceptance rule

ZLoop may ship only when all mandatory acceptance criteria pass independent verification and required review has no blocking finding. INCONCLUSIVE is never treated as success.