- Category
- AI Agents
- Rank
- No. 1044Tools index
Previous survey · No. 1024 ·
- Pricing
- Open Source
- Type
- AGENT
- Builder
- stanford-iris-lab
- GitHub
- 1.2k stars
- Date
About
Stanford IRIS Lab's agent scaffolding that hit 76.4% on Terminal-Bench 2.0 with Claude Opus 4.6.
What it does
Meta-Harness prepares a command-line agent before work begins. It records the sandbox’s directory, files, installed tools, package managers, and memory, then places that context in the first prompt. The agent controls a terminal through structured model tool calls and detects command completion by polling for shell markers.
Why it's ranked here
This is a focused benchmark artifact with a clear, testable idea. Preloading environment facts can remove repetitive discovery work, while native tool calls and early command-completion detection reduce avoidable agent overhead. The repository also provides a concrete Harbor invocation for reproducing its intended workload.
What's good
The scaffold attacks two practical sources of wasted time. Environment bootstrapping gives the model useful facts before its first action. Marker-based polling lets fast shell commands finish without consuming their full requested wait. Anthropic requests can also add ephemeral caching to the three most recent messages without modifying the originals.
Tradeoffs
The project targets Terminal-Bench and depends closely on Harbor’s Terminus framework, so it is not presented as a general agent platform. It requires Python 3.12 or newer and an external model API key. The README says details of the automated harness evolution are still forthcoming, limiting scrutiny of how the design was selected.
How to use it well
Use it when evaluating or extending autonomous command-line agents inside Linux sandboxes, especially when repeated environment discovery wastes early turns. Treat it as a scaffold to run through Harbor and inspect at code level. It does not provide a standalone user application, hosted service, or broad framework for non-terminal agent workflows.
Technical notes+
agent.py subclasses Harbor’s Terminus2 as AgentHarness, supplies native execute_commands, task_complete, and image_read tools, and uses tmux command markers for early completion detection. It also injects an environment snapshot before the loop, wraps blocking coroutines with a 600-second timeout, and loads prompt-templates/terminus-kira.txt. anthropic_caching.py deep-copies messages and applies ephemeral cache controls to the latest three messages for Anthropic or Claude model names. pyproject.toml declares Python >=3.12 plus anthropic, harbor>=0.1.44, litellm<1.82.7, and tenacity. Package metadata reports version 1.0.0, while AgentHarness.version returns 1.1.0.
Observed
- Primary language
- Python
- Runtime requirement
- Python 3.12 or newer
- Packaging
- Python project metadata declares anthropic, Harbor, LiteLLM, and Tenacity dependencies
- Interface
- Harbor agent import path with native terminal, completion, and image-reading tool calls
- Platform context
- Prompt targets command-line tasks in a Linux environment
- Framework base
- Extends Harbor Terminus-2 and Terminus-KIRA
Read from README.md, pyproject.toml, agent.py, anthropic_caching.py, prompt-templates/terminus-kira.txt.
What it can do
Execute terminal commands through AI agent scaffolding
Natural language instructions or commands → Terminal command execution results
Automate command-line workflows
Task descriptions and system environment → Completed terminal-based tasks
Provide AI agent framework for terminal operations
Agent configuration and task parameters → Structured agent responses and actions
Benchmark AI agent performance on terminal tasks
Terminal-Bench test cases → Performance scores and metrics
Scaffold Claude Opus model interactions
User queries and system context → Enhanced AI model responses
Intel on Meta-Harness
Tags
Tech Stack
Comments (0)
No comments yet
Editorially curated, with community endorsements as a secondary signal. Corrections welcome.