Twitxy Saved

Rahul

Verified account This account is verified. Learn more

@RHLSTHRM

CAIO @codos_ai

Prev: Head of AI @lifiprotocol / @everclearorg / @Tesla / @NASA. 3x founding CTO.
github.com/rhlsthrm Joined December 2009
3.1K Following 6.6K Followers 3.3K Posts

Latest posts 5

Rahul Verified account @RHLSTHRM
Software factories only feel temporary if you're already living in the future like Thariq. The companies we work with are still arguing about whether a $20 a month Copilot seat per engineer is too much.

For them the hard part isn't building the factory, it's getting the jobs, the access and the review loop clear enough that an agent can run them at all, and no model upgrade skips that step.

How far behind the frontier do you think the average engineering team is?
Thariq Verified account @trq212 23h
software factories feel like a temporary problem. the models will learn to build them for you

the more interesting question to me is how to build a workspace where agents can collaborate with you
View quoted post on X
Rahul Verified account @RHLSTHRM
2.5% to 41.3% of real freelance projects in 13 months is a wild jump, but look at what the agents were handed: a clear brief, a defined deliverable, and someone who already knew what good looked like.

Most teams don't have that for their own work, which is why the same models feel flaky inside a company and great on a benchmark.

Who on your team writes the best briefs, and are they the ones getting the most out of agents?
Chubby♨️ Verified account @kimmonismus 2d
Talking about acceleration: In 13 months, the best AI agent went from completing 2.5% of real freelance projects on the Remote Labor Index to 41.3% with Claude Fable 5.2.

These were paid projects that took human professionals 29 hours on average, with briefs provided and client calls excluded.

So yeah, white collar work will get a wake-up in 2027.
View quoted post on X
Rahul Verified account @RHLSTHRM
It took us hundreds of runs to notice one of our agent routines was busy doing nothing: it never failed and never changed anything, just quietly ran.

A cost cap would never have caught it because it was cheap, so we built the pause into the harness, and a routine that does nothing for 3 days now switches itself off and pings whoever owns it.

A failing agent gets noticed fast, but an idle one can run for weeks, so how would you even catch yours?
Rahul Verified account @RHLSTHRM
Been saying this for months: x.com/rhlsthrm/statu…

The key is an MCP server can’t just be prompts that anyone’s Claude can do, otherwise your MCP is just a skill. There has to be some type of proprietary data involved that’s the only moat.

Nicolas’s example of HR software’s hiring data is good. Hosting the database (ie as a CRM or task management software) is not enough of a moat. As models get better it gets easier to spin this up in one shot, fully customized to your use case. You need data the model doesn’t have, that’s all you can charge for.
Nicolas Bustamante Verified account @nicbstme 3d
I came to the conclusion that most SaaS should become an MCP server.

we’re getting the economics of AI software backwards. A lot of AI startups are building their own chat interfaces, agents, and workflows, essentially reselling tokens from the big labs at low to negative gross margins.

What if SaaS companies did the opposite?

Take HR software. Instead of building an AI assistant, become the best MCP server for HR. Expose tools to:

- Benchmark salaries against proprietary market data
- Run compensation reviews and calculate pay equity
- Model salary increases against a department’s budget
- Recommend promotions based on company policies
- Update employee compensation in the HRIS
- Manage approvals, permissions, and audit trails

Let ChatGPT, Claude, or Copilot handle the conversation, reasoning, and orchestration. Let the customer pay for the tokens.

The SaaS company keeps the valuable stuff: proprietary data, integrations, business logic, permissions, compliance, and the system of record.

You still charge a SaaS subscription, but you’re no longer paying to generate every answer. You can ship faster, with fewer engineers, and potentially restore software-like gross margins.

The next generation of SaaS might not have much of a UI at all. Just an incredibly good MCP server.
View quoted post on X
Rahul Verified account @RHLSTHRM
I assumed Slack would be the fast way to get approvals for our agents, since everyone lives in it all day.

The median wait in Slack turned out to be about 26 minutes, with almost half over an hour, against about 28 seconds in web chat. A notification is easy to ignore in a way a conversation isn't, so we're keeping approvals inside the conversation where the work is happening.

When your Slack is full of DMs from agents, you start to gloss over them. "Approval slop" is a real thing, effective human-in-the-loop is a UX exercise more than anything else.