Sarthak Rastogi, who writes a Substack newsletter on building AI agents, has released agent-engineer, a free open-source add-on meant to stop coding assistants from declaring their own work good without checking it. Rastogi built the tool himself, so this is a maker’s announcement, and the only evidence for it is a test he designed.

The add-on works with Claude Code, Codex CLI, Cursor, OpenCode, and Gemini CLI. It is a released tool, not a proposal: the code sits in a public GitHub repository under sarthakrastogi, and installation takes a clone plus two commands. In Claude Code those are a marketplace add and a plugin install. The post calls it “fully open source” but never names a licence, so check the repository before adopting it at work.

The package is counted as 11 skills, 5 hooks, and 4 subagents, which hides three plain ideas. Skills are markdown files of checklists and decision rules that load only when relevant. Hooks are one Python script that runs on session events. Subagents are separate helpers for jobs that would swamp the main conversation, such as reading 100 logs of agent runs.

The skills that matter are the ones aimed at the failures Rastogi says he hits daily. One covers testing: study real failures first, prefer plain code checks over AI judges, run each test several times, and gate releases in CI. One covers tool design for agents, pushing typed inputs, capped response sizes, and error messages the model can act on. A guardrails skill applies Simon Willison’s “lethal trifecta” (an agent that holds confidential data, reads content from strangers, and can send messages outward) and says to remove one leg. Others cover tracing and prompt caching, and one tells the assistant to write success criteria before choosing an architecture.

The hooks do the policing. They point the assistant to the right skill, stop it pasting API keys into the code, and ask for evidence before it says it is finished. That last check fires at most once per session and can be switched off. The post says the plugin collects no data and the hooks cannot reach the network. It only activates in projects that import an AI model library or agent framework.

Rastogi’s evidence is a blinded comparison on six agent tasks, three runs each, graded against a checklist he wrote before any run. Claude Code running Opus 5.5 met 53% of requirements with the plugin and 34% without. Two of the six tasks were written and run before he changed the plugin in response to results. The harness, rubrics, and fixtures are on a separate branch, and he estimates a full rerun costs $10 to $15 and takes about 30 minutes.

The sharpest example is a task asking whether a rewritten prompt was better. All three unassisted runs answered yes and predicted an accuracy gain, though nothing had been run. With the add-on installed, one of three did so. A plugin run also wrote a report claiming 100% accuracy from a keyword simulation, then corrected itself in the final message while leaving the false report on disk. Rastogi scored that run as a failure.

The limits are plain. That is eighteen runs per side, one model, and tasks written by the person selling the fix. Even with the plugin, nearly half the requirements went unmet. The “10x” in the post’s title has no number behind it; the only measured gain is the 19-point gap.

The post offers no stars, install counts, or contributor numbers. Its figure of 13,000 readers refers to an earlier explainer, not this tool. Rastogi is asking users to file issues and submit new test tasks, which reads as an early-stage project.

Teams building agents can check the claim cheaply: run his harness against your own repository for roughly $15 and see whether the gap survives contact with your tasks.

Based on Sarthak Rastogi’s post on his Substack newsletter, AI Agent Engineering (undated in the scrape).