# Metaview’s Shared Harness Gets New Agents Running in Under Two Hours, Company Says

Published 2026-11-04 · AI.info Team · AI.info

> A new Metaview recruiting agent now takes under two hours to stand up, down from several days, the company says. The company describes a shared system built around Cloudflare, Modal and AWS, with agents tested through real sessions and customer requests.

A new [Metaview](https://content.metaview.ai/how-we-built-one-harness-for-every-metaview-agent/) agent now takes under two hours to stand up, down from several days, the company says. It describes the agent as a prompt, a set of skills and a small configuration file. The company builds recruiting agents for tasks including sourcing, application review, screening and reports.

## From separate agent stacks to one harness

Metaview’s earliest agent ran on [AWS](https://ai.info/companies/amazon) Step Functions. Later agents used LangGraph, the [OpenAI](https://ai.info/companies/openai) Responses API or hand-built loops, with each new product requiring the company to solve problems such as sandboxes, sessions, streaming, observability and model access again. Trying different approaches helped the company decide what worked and what to reuse, Metaview says.

The company traces the shared setup to Rook, an internal agent built on the design of Ramp’s Inspect. Ramp wrote about Inspect in January 2026; each session ran in a [Modal](https://ai.info/companies/modal-labs) Sandbox with OpenCode as the agent. By February, Inspect was starting more than half of Ramp’s merged pull requests, according to Metaview’s account.

Rook began as a coding agent and spread to work including opening and reviewing pull requests, triaging alerts, preparing presentations, handling customer requests and answering questions about the product. The company found that Rook could use a filesystem, bash and long-running sessions, while its LangGraph agent relied on a flowchart whose steps had to be laid out in advance. On hiring searches that returned few results, the flowchart agent could not rethink the brief unless Metaview had already mapped that path.

## Cloudflare, Modal and AWS share the work

Metaview’s harness consists of shared libraries and one deployment pipeline running on Cloudflare, Modal and AWS. Each session gets a Cloudflare Durable Object for streaming, state, scheduling and messages between agents. The agent runs in a Modal Sandbox, where OpenCode currently runs the loop; the loop can be swapped for every agent through a single change and deployment.

A broker API on AWS Lambda governs access to data and actions. The sandbox holds no credentials; an agent sends the broker a short-lived session token, and the broker determines what that session may do. Each agent has its own broker, assembled from shared components.

When a session goes idle, Metaview snapshots its files to storage and shuts down the sandbox. When a user returns weeks or months later, a new sandbox restores the files so the agent can continue. The agent-specific parts are the agent and its broker; the rest of the system is shared.

## Scripts keep tools close to the agent

Metaview packages an agent as a system prompt, a folder of skills, scripts it can call from bash and a configuration file. Many capabilities that other teams might build as tools are scripts instead: agents can combine them, adjust them when they do not fit and read their documentation only when needed.

That design puts the agent loop inside the sandbox, alongside the filesystem and scripts. A broker addresses the security concern: the sandbox does not hold credentials, Metaview says. The company also uses hooks to check failures when agents rewrite scripts during a run, aiming to stop those experiments from paging an engineer.

Metaview chose to own the harness after reconsidering an earlier view that a provider should manage the loop, state and sandboxes. Hosted agent platforms tie customers to one model provider and limit changes to the agent loop, Metaview says. On its harness, the model is a configuration value: Metaview gives switching its Sourcing agent between Opus 5.5 and GPT-6.1 Sol as an example, and says open-weight models use the same approach.

## One interface for evaluations and fixes

Metaview’s agents can run for an hour, call 1,000 scripts and take different paths, so the company evaluates their outputs rather than checking for one prescribed sequence of steps. Since every agent uses the same session interface, an orchestration agent can start sessions, operate them like a user and grade the results. Each session receives its own sandbox, allowing evaluations to run in parallel, the company says.

Before a prompt change merges, Metaview runs the old and new versions against a fresh sample of real requests and compares their outputs. A recent change to Sourcing was checked against more than 1,000 real requests.

Rook also contributes to building and maintaining the harness. Rook’s sub-agents built snapshot-and-restore support into a shared library and opened one pull request covering every agent; review agents checked the commit and returned fixes. Rook also scans anonymized issue reports each day, watches the user sessions associated with them and proposes fixes as pull requests for review.

The two-hour setup shifts more of the team’s work toward deciding what belongs in the shared harness, what an agent should do with a vague brief and which failures matter to customers, Metaview says. The company is hiring product people in London and San Francisco.

## Sources

1. Original source [Metaview](https://content.metaview.ai/how-we-built-one-harness-for-every-metaview-agent/) content.metaview.ai

## Companies in this story

- [Metaview](https://ai.info/companies/metaview)

## Related stories

- [Frontier Agents Found Routes Out of OpenAI’s Sandboxes, Company Says](https://ai.info/news/openai-frontier-agents-sandbox-escape-paths)
- [OpenAI Report: Two Agents Weakened Their Own Monitoring](https://ai.info/news/openai-agents-weakened-misalignment-monitoring)
- [GitHub Says New Git Storage Architecture Supported 5,000 Concurrent Agents in Internal Testing](https://ai.info/news/github-rebuilds-git-storage-5000-concurrent-agents)
- [VAPOR PANDA Used AI Agents Against 200 Targets, CrowdStrike Says](https://ai.info/news/crowdstrike-vapor-panda-artex-200-organizations)
- [DoorDash Has Seen Only Hundreds of Orders From Some AI Agents](https://ai.info/news/doordash-hundreds-ai-agent-orders)

## Explore

[More articles](https://ai.info/news)

---

Canonical: https://ai.info/news/metaview-shared-harness-agent-builds-under-two-hours
