Zenith: A Continuous-Improvement Harness for Long-Running Tasks

Technical report from Intelligent Internet (2026) on how an agent harness should control work that may run for days or weeks, where the dominant failure mode is premature completion rather than inability to make progress.

Read the report (PDF)

Abstract

Long-running agents often fail not because they cannot make progress, but because they stop before the task is truly complete. We tested five harness designs across eight long-horizon tasks to isolate the control mechanisms that matter: repeated gap-finding, revisable planning, independent verification, adaptive orchestration, and stopping discipline.

RALPH is the strongest simple baseline because it forces each new session to reopen the gap between the current project state and the original requirement. But RALPH is expensive and has no principled stopping rule.

Our Zenith method keeps the useful parts of repeated review while making the loop adaptive: the orchestrator dynamically allocates workers, testers, reusable skills, replanning, and stopping decisions. In this study, Zenith achieved the best mean rank while using less than half of RALPH's per-task cost.

Installation

Copy and give this to your agent harness:

/goal Read the readme at https://github.com/Intelligent-Internet/zenith, detect if using claude code, codex or both, install requirements, install and run Zenith (ie uv run zenith etc like in the readme) and create a new skill called /zenith - when used (along with an additional prompt) it will call the skill: the minimum skill content should be: """First read .claude/orchestrator_prompt.md and treat it as your primary role, then use Zenith to run this mission.""" and afterwards you can add information about Zenith Harness, based on the readme and the technical report (inside the repo), and info on how to start zenith if its not already running. Change the skill to use .codex when using in codex. If both hernesses are available, make sure to add these skill to both of them correctly. Infact, there might be more harness options (hermes for example). See what is supported in §zenith/src/zenith_harness/providers.py, and for those that you detect are present, add their skills correctly. When Finished, confirm to me that Zenith is installed, running and ready, explain a bit about Zenith, and why and when to use it.

This will:

• install zenith with requirements
• start zenith using uv
• create claude code and codex skills to run & use zenith

Then, use /zenith like so - in claude code or codex, type:

/zenith <your instruction or query>

Or install it manually:

Zenith is a small MCP/ACP harness that runs a coding agent as a multi-agent orchestrator. See zenith/ for the full package.

Requirements

• Python 3.11+
• uv
• Node.js 22+ and npm
• Claude Code or Codex

Install

cd zenith
uv sync
uv run zenith --help

Install the ACP adapters globally for the agents you want Zenith to run:

# Claude workers/validators
npm install -g @agentclientprotocol/claude-agent-acp
command -v claude-agent-acp

# Codex workers/validators
npm install -g @agentclientprotocol/codex-acp
command -v codex-acp

Initialize a workspace

Initialize the project workspace Zenith should operate on. This is your target app/repo, not the Zenith source checkout:

# Claude Code, from this Zenith checkout
uv run zenith init --workspace-dir /path/to/your-app --agent claude

# Or Codex, from this Zenith checkout
uv run zenith init --workspace-dir /path/to/your-app --agent codex

Run a mission

Start your agent from the initialized project workspace:

cd /path/to/your-app

claude
# or
codex

Then ask the agent to read the generated orchestrator prompt:

First read .claude/orchestrator_prompt.md and treat it as your primary role, then use Zenith to run this mission.

<your instruction or query>

For Codex, use:

First read .codex/orchestrator_prompt.md and treat it as your primary role, then use Zenith to run this mission.

<your instruction or query>

Zenith

A single orchestrator session reads task state each turn and decides what to do next: spawn worker or tester subagents, register a reusable skill, replan, or stop. Workers and testers run in their own contexts and report back; the orchestrator integrates their results before the next decision.

Results

Frontier SWE Benchmark

On the Frontier SWE benchmark, Zenith — running on GPT-5.5 — ranks first overall, leading on implementation, performance, and dominance against frontier models paired with their native harnesses.

Ablation Study

To isolate the control mechanisms that matter, we compared Zenith against RALPH and three reduced harness variants across eight long-horizon tasks. Zenith achieves the best mean rank at less than half of RALPH's per-task cost.

_{A "win" is a task on which the method ranked first; the eight wins partition the eight benchmark tasks.}

Citation

@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}
}