Hey YC! 👋 We're Akshay and Ashwin, and we're building Medley.

You type /mission and give it the outcome, the budget, and the constraints. Medley builds a live dependency graph, coordinates Claude Code and Codex workers, reviews what comes back, and keeps going when execution reveals more work. It's free and runs on the subscriptions and API access you already have.

https://youtu.be/xe7bj3b5Aw0

❌ The Problem

Coding agents got good at bounded tasks. Once a job outgrows one session, you become the harness. You break the work into prompts, move context between agents, review outputs, and decide what happens next. There's no shared definition of done, so "finished" just means "looks finished."

🛠 Our Solution

/mission takes on that orchestration. Three things sit between your outcome and the models:

• A contract. Before any work starts, Medley sets the objective, the acceptance criteria, the constraints and budget, and the conditions that should make it hold or stop. That's what "done" gets measured against later.
• A live graph. Medley doesn't write one giant plan upfront or spawn a swarm by default. It plans the next useful set of steps, runs them, reviews what came back, and updates what should happen next. Each node carries a brief, its dependencies, and a verification target. Sometimes one worker should carry a tightly coupled job end to end; when work is genuinely independent, Medley runs it in parallel. Work is routed across Claude Code and Codex workers - and with your own OpenRouter key, Codex workers can run models like Kimi and GLM.
• An acceptance stage. When a batch lands, a separate reviewer role inspects the outputs and changed files, runs the contract's verify commands, and decides whether the evidence satisfies the contract - not whether the work merely looks finished. If it doesn't, the mission replans.

You can watch and steer the mission in your browser. Anything that needs your judgment or approval - publishing, sending outbound, spending, contacting people - lands in the attention queue. Everything else keeps moving. 

📊 Why This Works

The model is only the local capability. The harness decides where to apply it - the contract fixes the target, the graph splits the work, routing matches each task to a worker, and an independent acceptance stage runs the real verify commands and replans when the contract isn't met.

Before launch, we pressure-tested a benchmark-specialized version of the same architecture on four complete public benchmarks: 93.25% on Terminal-Bench 2.1, 71.37 on HealthBench Professional, 93.24% on ViBench, and 60.98% on DrugDiscoveryBench. Those are the highest reported results on each, and they span coding, healthcare, and drug discovery.

No new foundation model. The gains came from how the work was decomposed, routed, carried across contexts, and reviewed.

Full methodology and results are on our blog.

🙏 Our Ask

Try it. Medley is free on Mac. Install it from Claude Code:

/plugin marketplace add Spine-AI/medley
/plugin install medley
/reload-plugins

Then run /mission. medley.sh

Tell us what you think. We're shaping the product around how people actually use it. Email support@getspine.ai with what you tried, where it worked, and where it broke - and where you'd want Medley to keep going versus stop and ask.