hip-agent: A Harness Embedded in the Prompt (5-minute read)

Existing coding harnesses are designed for human users, making them difficult for agents to use as sub-agents. hip-agent is a 'prompt-embedded harness' designed to solve this, consisting of about 200 lines of Python and a single Codex API module.

Core Design Principles

1. Prompt-embedded: Instructs the model to read the harness's source code and understand its behavior, providing only two tools: sh and view_image.
2. Leveraging OS runtime: Configuration is handled via environment variables, behavior via shell commands, and sub-agents via child processes, reducing complexity.
3. Compliance with existing protocols: Follows the Agent Plugins specification and Claude Code's hook contract, and supports the Codex CLI session file format.

Benchmark Results In an evaluation of 113 tasks on DeepSWE, hip-agent showed performance comparable to Codex CLI 0.147.0.

• Resolution rate: hip-agent 64.6% (73/113) vs Codex CLI 63.7% (72/113)
• Number of model calls: hip-agent 187 vs Codex CLI 208
• Execution time: hip-agent 58 minutes vs Codex CLI 52 minutes

Implications and Limitations hip-agent originated from the concern that a fixed harness could become a bottleneck for improving model performance. Because the model can read and modify the harness (repairable), it can be used for training where models directly build harnesses tailored to specific tasks in the future. However, the results are based on a single run each, so there is no error margin, and hip-agent does not log token usage, so that metric was not compared.