Implementing Autonomous Agent Loops for Compounding Workflows

The Shift to Loop Engineering

Loop engineering moves beyond simple task completion by orchestrating multi-session agent workflows that persist state across time. Instead of relying on a single prompt to finish a task, developers build systems where agents trigger each other via shared file systems, allowing for autonomous, cross-session work that compounds over time.

Core Components of an Agentic Harness

To enable autonomous work, the codebase must be structured as an agentic harness that is legible, executable, and verifiable.

• Legible Codebase: Maintain an agents.md file (roughly 100 lines) that acts as an index for documentation and system rules. Use programmatic linting to enforce constraints, such as preventing imports from legacy folders, to reduce reliance on the agent's internal knowledge.
• Executable Environment: Ensure the agent can spin up a local development server without manual intervention. Use work-tree friendly setups so multiple parallel agents can test changes in isolation without conflicting.
• Verifiable Output: Provide agents with tools like Playwright to perform end-to-end tests and record video clips of the results. Crucially, do not allow agents to self-verify; instead, spawn a separate, read-only verifier agent to review the work against a defined PR checklist.

Shared Memory and Artifact Systems

Compounding effects are achieved by using a shared file system as a "brain" where agents read and write artifacts.

• Artifacts: Define specific folders for different outputs (e.g., signals, docs, tasks). Each artifact folder should contain a README defining the schema, process for additions, and metadata structure.
• Loop Contracts: Every loop requires a README acting as a contract. This file must define the loop's goal, workflow, backlog, and a timeline of past actions so the agent understands its state before beginning new work.
• Global Logs: Maintain a global_work_log.md where agents record major actions. Before starting a new task, agents read the last 5 to 10 entries to maintain context across different domains.

Compounding Workflows

By connecting loops, the output of one agent becomes the input for another. For example, a support loop identifies product frictions and logs them as signals. A separate product growth loop reads these signals to prioritize features, while an engineering loop monitors the same signals to automatically implement bug fixes. This creates a self-improving system where human intervention is only required for high-level review.