AI NewsDev toolsAnnouncement

Reason Machines launches an agent harness that runs in Slack

Reason Machines launched on Y Combinator with an agent harness that runs on cloud VMs and inside Slack. It reports 20,000 users and a benchmark result against Codex.

AI News

Editorial2 min read

LinkedInX
Reason Machines agent harness launch page on Y Combinator

Why it mattersA team paying for Codex, Cursor or Devin has a new option with a published number against them. The option runs on the team's own API keys and subscriptions.

A team paying a flat fee for a coding agent has not had many ways to lower the bill. Reason Machines wants to change that by letting the team bring its own keys.

Reason Machines launched on Y Combinator's launch page. Its YC batch is Fall 2026. The pitch: an agent harness that runs on cloud VMs, uses the team's own API keys and subscriptions, and lives in Slack. The company reports 20,000 users, 14,000 X followers, and 3 million impressions since launching four weeks ago.

What it does

The product is a harness, which is the layer that drives a model through its tool calls. Reason's harness runs on cloud Linux and Mac machines, local machines, or SSH boxes. A session can move across those machines without losing its state. The company says a team can run hundreds of sessions in parallel.

The harness takes the team's existing API keys or coding subscriptions. The company names Grok, Copilot and Kimi subscriptions. It uses auto-failover to keep going if one key hits its limit. A user mentions @Reason in a Slack thread and gets back a pull request, a slide deck, or an answer.

The company's own benchmark

Reason Machines says its minimal design wins against larger harnesses. The company claims a 30 to 80 percent lower cost per task than Codex, Cursor or Devin. The figure comes from its own benchmarking. Nothing independent has tested the claim.

Reason reports one specific comparison. On the Coding Agent Index, its harness scored 65.1 against Codex's 62.5. The company says that was at 22 percent lower cost per solved task. Reason's own blog post contains the full write-up. A reader should treat the figure as Reason's own number, until a third party runs the same test.

The founders

The company's launch page says Adi and Sven studied EECS at UC Berkeley. It says Adi worked in applied machine learning at Scale AI, and previously built a Rust-based IDE. It says Sven built AI models at a Fortune 100 company. The page names no customer outside its own user count.

What a buyer should check

The product's claim rests on the harness design and on the team bringing its own keys. A buyer paying for Codex would need to confirm the cost claim against its own workload. The 20,000-user figure is Reason's own count after four weeks. It is not a revenue number, and the launch page gives no retention figure. YC companies get 1,000 US dollars in product credits through a YC deal, which the page states.

A team that cares about auditability should ask what runs where. The harness runs on cloud VMs that the user signs into. The model call still goes to whichever provider the user's key points at.

Source

This item was written by an AI system from the linked source. Reveneau is responsible for what it publishes.

Share
LinkedInX
Start a project