Fixing AI Agent Laziness with Unlazy

The Breakthrough

Unlazy solves AI agent laziness by replacing self-reported task completion with a mandatory 'gates' ledger, which forces the agent to execute verification commands and provide evidence for every sub-task before it is marked as done.

How It Works

• Task Decomposition: The skill recursively splits large requests into smaller sub-tasks, forming a tree structure. Each leaf node represents a discrete unit of work estimated to take at least ten minutes.
• The Gates Ledger: The system generates a gates.md file containing a checklist of outcomes. Each gate requires a specific command to prove success, the expected output string, and an evidence field that is populated only after the verification command runs.
• Orchestrated Execution: When the depth parameter is set to 4 or higher, the system switches to orchestrated mode, where a main agent manages sub-agents. Each sub-agent receives only the plan and its specific gates file, preventing context window bloat and focusing attention on the immediate task.
• Parallelization Fix: The default implementation processes tasks sequentially, which causes significant latency. By modifying the skill instructions to enable parallel agent execution, the author reduced build time for a full application from four hours to two hours using ten concurrent agents.

Context

AI agents frequently suffer from 'laziness' due to context window saturation, where they either report tasks as finished prematurely or omit difficult components of a job. Previous attempts to mitigate this, such as the Ralph loop or secondary model judging, failed because they relied on the agent's own summary or a separate model's interpretation of the conversation rather than objective verification of the work product. Unlazy addresses this by treating the agent's output as untrusted until verified by the system's own checker.

Notable Quotes

• "It doesn't tell you the agent is done, it proves it."
• "A ticked box with 'pending' still under it means the agent ticked that box itself, which is just the agent telling you it's done all over again."