Labbench

Labbench — an AI research agency, in the cloud

An AI Research Consultancy in the Cloud

Your research department, on retainer.

Point Labbench at your repo. It reads the latest papers, runs real experiments in cloud sandboxes, and ships pull requests — with a report and a podcast.

Scroll to see how it works

The problem

You shouldn't have to quit your job to keep up.

The running joke is that keeping up with AI research is a full-time job. Every week ships results that could change your product — and you're busy shipping the product.

How it works

Hermes runs the engagement.

A president agent orchestrates four stages, end to end. You review the PR.

1

Watercooler

Ingests your repo into a dossier: what you're building, where it's weak.

2

Library

Searches the literature across Linkup and Exa; ranks pitches against your dossier.

3

Lab

Writes real code and runs 30-day benchmarks in Cloudflare sandboxes.

4

Conference

Opens PRs against your repo, with an HTML report and a podcast.

On retainer. Hermes keeps your research and repo in a three-layer memory across engagements — and between runs it dreams: consolidating what it learned and forging reusable skills, so every engagement starts smarter than the last.

The product

Watch it work, live.

Five views into a running engagement.

Bench

Mission control: every agent, every dollar, live.

Watercooler

The dossier taking shape as your repo is read.

Library

Pitches ranked as papers come in.

Lab

Benchmark curves streaming from the sandboxes.

Conference

The PRs, the report, the podcast.

What we tested

The benchmark: Vending-Bench

Vending-Bench is a long-horizon test: an LLM runs a vending-machine business for many simulated days. Over a long run, agents lose the plot and go broke — the score is Total Assets at the end. Labbench read the research, wrote and tested real code on a fork, and shipped the change that kept the agent solvent: +49% more Total Assets.

Proof

Three hypotheses tested. Two PRs shipped.

The system is real end to end: Labbench read the papers, wrote the code, and ran the benchmarks in cloud sandboxes. The winning hypothesis — a typed business-state briefing from KV memory before every tool-loop request — lifted long-horizon Total Assets by +49% on this run, and Labbench opened the pull request.

Total Assets — this engagement's run

Vending-Bench fork · same in-run baseline · higher is better

$1,000 $500 $0 $634.50 Baseline $732.55 +15% PR #4 $949.00 +49% PR #5 · winner

Result of this recorded run vs its baseline. Long-horizon benchmarks are high-variance; across repeated runs the winner averages ~+27–38% (n=3). We re-ran to confirm.

The pull requests

watson-vending-bench · PR #5 · winner
typed business-state briefing from KV memory
$634.50$949.00  +49% The hypothesis: Before every decision, inject a compact typed summary of the business state (cash, inventory, prices, orders) pulled from the agent's key-value memory — so it never loses track of where things stand over a long run.
watson-vending-bench · PR #4
Reflexion — persist failure lessons across the run
$634.50$732.55  +15% The hypothesis: After a costly mistake, write a short lesson to memory and resurface it when a similar decision comes up again — so the agent stops repeating the same errors.

Built on

Real infrastructure, no mocks.

Cloudflare Workers Sandbox SDK Durable Objects Convex OpenAI GPT-5.6 Linkup Exa ElevenLabs

Technologies we run on

See it run.

A live engagement, end to end, in the cloud.

See the live demo →