Intelligent-Internet/zenith

Zenith – A Harness for Long‑Running AI Coding Agents

What it is – Zenith is a Python‑based agent harness that lets large‑language‑model coding agents (Claude Code, Codex, or Hermes) keep working on a project for days or weeks without stopping prematurely. It does this by running an orchestrator that repeatedly looks at the current state of the task, then decides whether to:

• spawn worker sub‑agents to write code,
• spawn tester sub‑agents to verify the output,
• register reusable “skills” for later reuse,
• re‑plan the remaining work, or
• stop when the job is truly finished.

Why it matters – In long‑horizon software‑engineering tasks the biggest failure mode is early termination: the agent thinks it is done even though the specification isn’t satisfied. Zenith’s adaptive loop (inspired by the RALPH baseline) keeps the process alive, re‑examining gaps and only stopping when a disciplined rule says the work is complete. In the authors’ 2026 technical report, Zenith out‑performed five other harness designs on eight benchmark tasks, achieving the best average rank while using less than half the cost of the strongest baseline.

---

Quick‑start

Option 1 – Let the LLM install it for you

Copy the prompt shown in the README into Claude Code, Codex, or Hermes. The prompt will:

1. Detect which coding agent you are using.
2. Install the Python dependencies with uv.
3. Install the required ACP adapters (@agentclientprotocol/claude-agent-acp or @agentclientprotocol/codex-acp).
4. Register a /zenith skill that forwards any subsequent instruction to the harness.

After the skill is created you can simply type:

/zenith <your instruction>

and Zenith will take over the task.

Option 2 – Manual install

# clone the repo and cd into the package
cd zenith
# install Python deps (uv required) and the CLI
uv sync
# check the CLI
uv run zenith --help

# install the ACP adapters for the agents you plan to use
npm install -g @agentclientprotocol/claude-agent-acp   # for Claude Code
npm install -g @agentclientprotocol/codex-agent-acp   # for Codex

Create a workspace that points at the target codebase you want the agents to work on:

uv run zenith init --workspace-dir /path/to/your-app --agent claude   # or codex

Then start the chosen agent (e.g. claude or codex) in that workspace and issue the orchestrator prompt as described in the README.

---

How Zenith works (high‑level)

1. Orchestrator session – a single LLM loop that reads a snapshot of the project (files, test results, etc.) each turn.
2. Decision step – based on the snapshot it chooses one of four actions: spawn a worker (writes code), spawn a tester (runs tests/validation), register a reusable skill, re‑plan the remaining steps, or stop.
3. Sub‑agents – workers and testers run in isolated contexts (via the MCP/ACP protocol) and report their results back to the orchestrator.
4. Integration – the orchestrator merges the sub‑agent outputs, updates the project state, and repeats.
5. Adaptive allocation – the orchestrator learns when to allocate more workers, when to focus on verification, and when to halt, reducing wasted API calls.

---

Reported performance

Benchmark	Model	Harness	Avg. rank (lower = better)	Dominance	Cost per task
Frontier SWE	GPT‑5.5	Zenith	2.06	92 %	$175.68
(others)	…	…	…	…	…

Ablation studies on eight long‑horizon tasks show:

• Mean rank: Zenith 1.38 vs. RALPH 1.75
• Mean cost: $175.68 vs. RALPH $407.58
• Wins: Zenith wins 5/8 tasks, the most of any method.

---

Repository layout

• zenith/ – the Python package (zenith-harness), CLI entry point, MCP server, bundled prompts/skills, and tests.
• technical_report/ – PDF, LaTeX source, and figures for the From RALPH to Zenith technical report.

---

Who should use it

• AI‑augmented software engineers who want a hands‑off way to let a coding LLM finish a large feature or refactor over many days.
• Researchers exploring control mechanisms for long‑running agents (gap‑finding, adaptive orchestration, stopping rules).
• Tool builders who need a reusable harness that can plug into Claude Code, Codex, or Hermes via the ACP protocol.

---

License

• Code: Apache 2.0
• Report & figures: Creative Commons BY 4.0

---

Citation (as provided in the repo):

@techreport{ii2026zenith,
  title       = {From RALPH to Zenith: Designing Harnesses for Long-Running Agents},
  author      = {{Intelligent Internet}},
  institution = {Intelligent Internet},
  year        = {2026},
  type        = {Technical Report},
  url         = {https://github.com/Intelligent-Internet/zenith}
}