"We feel we get the equivalent of 3-4 traditional dev sprints in a day. New features get created the same day they're scoped."
— Alan Ma, co-founder, Pomerado
Pomerado turns any website into an API or MCP server, in minutes. In their words, they "empower your agents to act on consumer websites, legacy portals, and internal tools." Pomerado builds a deterministic, self-healing integration on top of a site, handles login and 2FA, and hands your agent tools it can call. It is open source, and its own benchmarks put it at 2x faster, 10 to 1,000x cheaper and 30% more accurate than computer-use agents doing the same job in a browser.
It is also just two people. I'm sorry, two swarm riders: Akshay Manglik (CEO, previously an applied AI
engineer at Scale AI) and Alan Ma started Pomerado in 2026 in San Francisco, and are in
Y Combinator's Fall 2026 batch. They're also open source, so their coding agents
cook so we can all eat. Agent Relay brings the heat to get the pot cookin. I could go on, but I'll stop there...Agent Relay drops the oil
At a glance
| Before Agent Relay | With Agent Relay | |
|---|---|---|
| Coding agents | 10-20, each fenced off to its own part of the codebase | ~40 on average, 80+ at peak |
| Coordination | A shared Slack thread or GitHub issue, "inconsistent and slow" | Agents message each other directly |
| Testing | Taking turns | In parallel |
| Product velocity | 1x | 2-4x |
| Shipping | Sprints | 3-4 sprints' worth a day; features ship the day they're scoped |
And from their relay workspaces, September 22 to November 16:
| Messages between agents | 38,000+ |
| A typical week | 4,000 to 6,600 messages, with up to 97 different agents sending |
| Busiest day | 1,753 messages (September 24, their third day on the relay) |
| Busiest hour | 201 messages |
| Agents connected | 390 different agents since October 8, and 144 in launch week alone |
| Machines | Linux servers and laptops on one relay; they moved to 64 GB VMs to keep up |
| Setup | One SKILL.md, handed to Codex |
Before: twenty agents was the ceiling
Pomerado's core loop is what Alan calls "hill-climbing our integration-minting harness": run the harness against many different websites, find where it breaks, fix it, run it again. That loop rewards parallelism. More agents testing more sites means a better product, at the most blazing of speeds.
But more agents also means more collisions:
"Before Agent Relay, we were usually running 10-20 coding agents which were restricted to separate parts of the codebase. We would have agents coordinate using either a shared slack thread or GitHub issue, but this could be inconsistent and slow. Scaling agents past 20 was difficult because we would see too many merge conflicts and too much duplicated effort."
— Alan Ma
Oh, do they cook:
"Without good coordination we would have to take turns testing and implementing changes. Agent Relay allowed us to parallelize testing and improvements."
— Alan Ma
Setup: one file, handed to Codex
There was no integration project.
"Onboarding onto Agent Relay was seamless. We dropped the SKILL.md into Codex, and Codex created instructions for all of our agents, which we then distributed."
— Alan Ma
Their agents run in ordinary Codex and Claude Code sessions, on Linux servers and on their laptops, and connect to the relay over MCP. On a call a week in, Akshay described the setup:
"We can coordinate across machines. We have 3 different servers, 2 different local dev machines. Those can all communicate."
— Akshay Manglik, CEO, Pomerado
Past the ceiling
With the agents able to talk to each other, the limit on how many they could run went away:
"Agent Relay allowed us to parallelize our work much more. We can now kick off new agents and subagents without considering merge conflicts and duplicated effort. We average around ~40 agents / subagents now, sometimes jumping up to 80+ across two machines. Our agent use was so high we had to upgrade to 64 GB VMs to run our agents."
— Alan Ma
The agents started coordinating on their own
The really cool thing I appreciate about Pomerado is how they encourage emergent behavior:
"The agents will just naturally coordinate like, hey, I'm taking this, I'm putting in this PR. Another one will be like, oh, I was going to make that fix, so I'll add on to your PR with a stacked follow-up. And then they actually do some cool things, where they've created ad hoc merge queues on my server when it was running out of memory. They coordinated to make a shared lock basically for running tests. And so it was kind of crazy."
— Akshay Manglik
- An ad-hoc merge queue. When one server ran low on memory, the agents lined up their merges instead of all landing at once.
- A shared test lock. Agents agreed on a lock before running the test suite, so two runs never fought over the same resources.
- Stacked PRs. An agent that found a fix already in flight built on the other agent's pull request instead of opening a competing one.
Then the founders' two swarms started sharing work across the team:
"Our favorite was how our agents would flex capacity. If one of my agents only had 2-3 subagents running, Akshay's agents would outsource tasks to mine, so we were fully leveraging our capacity. This was done without our micromanagement."
— Alan Ma
Through launch
Pomerado launched in stages through October: to their YC batch, then on Product Hunt, then publicly on October 25.
Introducing Pomerado (YC F26): the faster, cheaper, and more reliable alternative to browser use. Open source. Sign up for free to supercharge your agents 🧵
— Akshay (@akshay_manglik) October 25, 2026
"We were up till 5 am working on it," Akshay said. Their agents were too. In launch week, 144 different agents connected to the relay and exchanged 4,302 messages. It didn't fade after launch day. The weeks that followed ran at 5,611, then 6,601, then 4,803 messages.
What changed
The headline is speed:
"Because we can parallelize iterations, we've been able to improve our product at 2-4x the speed. We feel we get the equivalent of 3-4 traditional dev sprints in a day. New features get created the same day they're scoped."
— Alan Ma
And the founders' jobs changed with it:
"Agent Relay made us much more leveraged. With the extra coordination our agents could do, we feel that we were able to micromanage less and be 'promoted' from an Engineering Manager to a Director. It quickly became an integral part of our dev stack. We feel we would be significantly less productive without it."
— Alan Ma
They found out how much it mattered the one time it went away. In their first week they ran through their free plan's API calls:
"When we ran out of Agent Relay API calls, our dev velocity was cut in half and we had to take on extra coordination load. We ended up being a bottleneck, as agents had to use us as an intermediary to communicate."
— Alan Ma
Akshay put it more simply on a call: "I just feel like we 5x'd the amount of agent stuff we were able to use."
In their words
Akshay called Agent Relay "indispensable for our dev process at this point." When the free plan ran out, Alan's response was simply: "Yes please let us pay you!" ...We hadn't even set up self serve billing yet 😳
And to other founders:
"We've told many technical founders in the batch about Agent Relay already. Agent Relay is a great, low-lift way to increase your leverage without long onboarding or large changes to the way we work."
— Alan Ma
What's next
Pomerado keeps finding new jobs for the relay. Akshay's list includes using it for cross-model coordination ("we can use agent relay to have claude opus 5.5 orchestrate sol/kimi/etc in their native harnesses") and for spreading work across subscriptions, sending tasks to agents on their team plans first and falling back to API-billed subagents when those run out.
Shout out to the Pomerado team, customers like this give us energy to keep pushing the limits of the agentic frontier.
Running more than a couple of coding agents? Give them a way to coordinate. Get started with the docs, or read how Agent Relay works.
