Ashwin V Raman (@AshwinVRaman) header
Ashwin V Raman (@AshwinVRaman) profile picture

Building performant long running agents at @Spine_AI (YC). ex-@Nvidia, @Amazon, @Instacart

Joined November 2013
Finally great to see the major labs following Google’s lead in pacing the frontier 🫡
OpenAI: “We’ve solved math” Anthropic: “Our AI is so powerful it’s going to kill you” Google: “Introducing Gemini 3.9 Flash! It’s 30% faster and 15% worse than the last Gemini”
2
99
Ashwin V Raman retweeted
The harness is the capability multiplier. Models are locally capable. They can reason, code, search, and use tools on the step in front of them. But ambitious outcomes are global: they depend on the plan, the handoffs, the context that survives them, which model handles each task, and what gets verified before the system is allowed to stop. We built Medley for that layer. The 'mission' harness has three core components: 1. a planner agent; 2. a contract-driven DAG of workers, where each node can use a different model/harness pair and its own instructions; and 3. a review agent that tests the assembled result against the contract and triggers focused repair work when it falls short. A separate benchmark meta-harness learns the benchmark-specific instructions. It selects unlabeled tasks, runs the mission harness, studies its traces, and edits the responsible planning, worker, review, or runtime skill. The adaptation is unsupervised and grade-blind: it receives no solved examples, labels, or grader feedback. The reason for the graph is simple: you cannot fully plan long-horizon work in advance. The information needed to solve it emerges during execution. You need an operating system that can update the graph without losing the contract. The resulting systems set new reported highs across terminal work, clinician conversations, full-stack applications, and drug discovery. Full report: medley.sh/blog/the-harness-i…
4
8
1
28
4,723
Claude Code and Codex are powerful. Once a job spans sessions, you become the orchestrator – splitting it up, moving context, reviewing results, deciding what happens next. Give Medley the outcome. /mission takes it from there. SOTA on 4 benchmarks.
Today, we’re launching /mission by Medley. Give Claude Code an outcome, not a task. Medley turns it into a live graph, coordinates Claude Code + Codex workers, then reviews and keeps going. SOTA on 4 public benchmarks—from coding to healthcare and drug discovery. medley.sh/
1
3
123
There are only three ways to get something real done as a founder. All three come with a cost. Say I need an outcome. "Find our next engineering hire" or "Optimize our onboarding for higher product retention." My options: 1. Do it myself and prioritize it over dozens of other outstanding tasks. 2. Hire someone - money and months we don't have yet. 3. Wire up some tools to limp through a piece of it. What's not on the list: hand it to AI and walk away. For a real, multi-step mission, that's still not an option. We're launching Medley soon - the platform where you hand AI a real role - set the target, the budget, and the deadline, and let it own the mission the way a teammate would. Join our waitlist at medley.sh?utm_source=x.
1
6
122
Ashwin V Raman retweeted
Spine beats Claude on outputs 89% of the time. That means less time spent on revisions and fact-checking results. See all of the ways users are leveraging Spine's state of the art deep research to generate client ready artifacts.👇️ youtu.be/vRzxZcv_Oas
1
1
7
118
Your agents and apps can now have easy access to the highest quality deep research in the market.
Spine API is here 🚀 You can now implement the state of the art into your own applications. Spine orchestrates parallel agents across 300+ models, runs deep research, and returns client-ready slides, docs, spreadsheets, images, and reports. Start with $10 in free credits, on us. Get started directly at getspine.ai/api
1
4
134
Anthropic’s latest release may have flown under the radar: their addictive new game, “Predict the Next Color.” Kudos to Anthropic for making probably the most visited status page on the internet weirdly entertaining.
2
43
Ashwin V Raman retweeted
We're announcing Teams and flexible pricing today. Teams: Add your whole team to one workspace. No extra cost per person. Everyone shares credits under one subscription. Pricing: Choose how many credits you want each month. Price adjusts. Go up or down anytime. You only pay for what you use. Already on a plan? Nothing changes unless you want it to.
1
1
3
110
I stopped checking Slack first thing in the morning. I have Spine do it for me. Today we launched integrations and automations on @Spine_AI. Agents use your tools when they need them. Spine handles auth and re-runs canvases on schedule with fresh data. A few other canvases we run internally: → Morning meeting prep from Calendar + Gmail + web research, saved to Drive → Weekly competitive intel from web + Slack, saved to Drive → Friday changelog from GitHub commits, saved to Notion Live on Product Hunt → tinyurl.com/55fcu4vr
1
23
We just made two changes to how Spine works. First, you can now add your whole team to your Spine workspace. No extra cost per person. Everyone shares one pool of credits under one subscription. Second, you're no longer locked into a fixed plan. You choose how many credits you want each month. Price adjusts to match. Busy month? Get more. Slow month? Get less. Changes happen right away and you only pay for what you use. We built it this way because the old setup was annoying. You'd either pay for credits you didn't touch, or run out halfway through a project. Neither made sense. If you're already on a paid plan, nothing changes. Same price. Same credits. You just have more options now.
1
1
33
San Francisco isn't the only place you need to be if you're building AI. I used to joke that if you were ever stuck on a technical challenge in SF, you just needed to open your window and shout it out and someone on the street would probably be able to help you. SF has a high density of AI builders and so much of what matters in AI gets shared in corridors and meeting rooms long before it shows up in a blog post or podcast. But there's another place where that happens: X. It’s hard to find another place where your curated trending news is about agentic memory and the theater of software compliance. If you're building in AI and not on X, you're missing signal that matters. Even if you're in SF. HackerNews gets an honorable mention but I prefer the personalized curation of X on most days. What other platforms are you all using to keep up?
1
4
94
Your AI team can now do your work for you while you sleep. Prompt. Set interval. Schedule. Fresh intelligence every morning. This is the future of managing AI agents 📷
Agent Automations is live 🚀 Schedule a swarm of AI agents to run on repeat. Daily competitor reports. Weekly trend digests. Real-time market alerts. Just give Spine a prompt and agents will work while you sleep. This is what an AI workforce actually looks like. See comments to get started for free.
1
1
4
304
Ashwin V Raman retweeted
Looking to make the most of Spine? Here's how in under 8 minutes! youtu.be/VM4bq2l67Ww
1
1
199
84% of the world has never used AI. 16% have used a free chatbot. 0.3% pay $20/month. 0.04% use coding tools like Cursor or Claude Code. This graph puts it in perspective. Each dot is 3.2 million people. (Credit: Ben Tann via Linkedin) The AI gap is real, and it's wider than most people think. If you're in tech or in SF, it's easy to assume everyone knows what AI can do, but the sobering reality is that most people have yet to see the current capabilities of AI. Most of that 16% tried ChatGPT for Q&A, got a mediocre answer, and never came back. They never saw what happens when you give AI a real task and let it run. AI has moved well past Q&A, and if that's all you've tried, you're missing what's already possible. The gap between free AI and what's available today isn't price, it's capability. Free chatbots give you answers, but the tools available now give you finished work. The models are getting better faster than founders can build products around them. Writing code is starting to get automated and the rest of knowledge work is next. That's why we're building Spine Swarms. Long-running agents that work while you sleep. No learning curve, no prompt engineering. You describe the project, like "research these 50 companies and build me a competitive analysis," and agents execute it. You get a finished deliverable, not a chat response. Try it at getspine.ai. Don't ask it a question, give it a task.
1
135
I’ve been frustrated by a massive gap in my workflow. When I’m coding in Cursor or using Claude Code, I have a high-velocity environment where I can describe the task and get code almost ready to ship. But the moment I switch to any other work (strategy docs, marketing assets, prospecting and making customer decks) it doesn’t feel the same. They seem impressive at organizing files and folders but I’m stuck in linear chat windows and working with local files where I can’t easily inspect the work, edit artifacts directly, or run experiments in parallel. I’m building Spine Swarm to solve this. It combines the power of AI agents with an IDE for general knowledge work. Instead of just chatting, it’s a visual workspace designed for knowledge work: 1. Visual organization & collaboration: See the artifacts being created by the different AI agents, inspect them, and edit them in place. 2. Parallelization: Run multiple variations of a task simultaneously to compare outputs instantly. 3. Model agnosticism: We route between 300+ models, automatically picking the best one for the specific task. 4. Zero setup: All the power of a complex agentic workflow, but accessible immediately in the browser without complex setup or configs. Dispatch your first agent swarm for free👇
1
1
1
2
159
Everyone's talking about Claude Code and OpenClaw. The takeaway is clear — AI agents that run autonomously and finish real work are the future. But it shouldn't require a terminal to use them. Spine is live on Product Hunt, hunted by @garrytan PH Link: producthunt.com/products/spi…
1
6
76
Time to put your AI team to work. 7,000+ people have been waiting to get their hands on Spine Swarms. Today it's open to everyone — no waitlist, no invite. State of the art reasoning and research that translates to polished deliverables — docs, decks, dashboards, spreadsheets, and more. Link in the comments. Try Spine Swarms: tinyurl.com/46a43dwx
1
2
7
219
Our team of 8 — 5 engineers — just beat Google, Perplexity, and OpenAI on the top research benchmarks. Here's how architecture beat scale. 🧵
1
37
7/ Methodology: → GAIA: validation set — consistent with Manus, OpenAI, Genspark. SOTA as-is. 76% after correcting mislabeled questions. More on that later. → DSQA: fully correct on the entire 900-sample dataset. Not a subset. Drop a comment for the full dataset runs. Every step is on the canvas.
1
48