adk-loop-lab
A documentation site for building bounded, inspectable agent loops with Google ADK.
A useful agentic loop combines durable state, selective memory, bounded execution, verification, explicit progress tracking, and deterministic stopping conditions.
The Loop Lifecycle
The project does not treat the model call as the system. The model is one step inside a loop controller that reconstructs context, executes bounded work, verifies results, records evidence, and then decides whether another iteration is justified.
Three Planes
- Control plane: deterministic lifecycle ordering, checkpoints, stopping policy, and budget enforcement.
- Execution plane: ADK agents, Gemini calls, workflow composition, and sandboxed tools.
- Data plane: SQLite-backed state and memory, plus JSONL event traces for auditability.
Three Examples
Bounded Document Refinement
A generator-critic loop with hard checks for word count, examples, and unsupported citation phrasing.
Level 2Evidence-Driven Research
A research tracker that turns questions, claims, and source IDs into a claim-evidence matrix and a final report.
Level 3Resumable Coding Loop
A coding workflow with observer, planner, implementer, sandboxed execution, and failure-memory guards.
Quick Start
git clone https://github.com/rmax-ai/adk-loop-lab.git
cd adk-loop-lab
uv sync
uv run adk-loop run level-1For a deeper read, continue with the architecture page, the concepts guide, or the examples index.