Harsh Raj

Co-founder & CTO @UseAgentrail · Agent conversion infrastructure for developer platforms · EIR @joinef · sharing the experiments and failures

[Agentrail @UseAgentrail]

spent years getting google to notice us, only for an AI agent to get stuck on our signup page 🫠

[Nikunj Kothari @nikunj]

If you are visiting SF, and trying to meet more people, here are a random list of things you should remove from your vocabulary..

> swap notes
> "synergies"
> pick your brain
> networking
> "san fran" (tourist tell)
> grab coffee if you're free

If you want to meet interesting people here, show your work. The most interesting people I know in the city are not attending that fancy large conference - they are probably at their office.

So what's the work? It can be as simple as a thoughtful cold email, a five line pointed DM, or complex as some side projects that you have shipped, or a meetup that you are organizing.

If you put in the effort, it shows, yes, even in this AI era. I still get so much slop across the board, that if I see anyone put in the effort, I try to go out of my way to meet them.

Even the busiest people have 15 mins to give you, if you give them a reason to!

[Harsh Raj @h_v_raj]

the missing piece is knowing where an agent got stuck, not just what it did. a trace of 12 tool calls won't tell you whether auth failed, context got lost, or the product gave it no clear next step. that's what we're researching at https://t.co/lP6P7fl5AW

[ghxst @gitghxst]

observability for agents is going to be one of the biggest categories in software.

Today, when an agent runs for 40 minutes, calls 12 tools, spends $30 in tokens and changes 9 files, all you usually get back is "done ✅".

That's fine for small tasks. It's not fine when agents touch production, customer data and money.

We built a whole industry to monitor servers. Agents need the same thing, but harder, because they make decisions instead of just running code.

[Harsh Raj @h_v_raj]

the missing layer is deciding whether an agent should make a request at all: who authorized it, what it can access, and how to trace what happened when it goes wrong. that's the choke point between agents and every product they touch.

[Karan Vaidya @KaranVaidya6]

The agent cloud will be 100x bigger than the classic cloud. But one of its most important layers is missing from this list.

Cloudflare became a giant by sitting between humans and applications, securing and optimizing every request.

The next Cloudflare will sit between agents and applications.

Except agents don't just fetch information. They take actions, make decisions, hit permission boundaries, and encounter attacks.

The infrastructure that sees these interactions can learn what works, what fails, and what's dangerous across millions of agents.

[Harsh Raj @h_v_raj]

@ShubhamInTech @Superconductor @SuryaRajendhran

[Harsh Raj @h_v_raj]

@crucianijm @DavidKPiano not a clear pattern yet. that's the problem: when an agent gets stuck, it usually retries or takes another path without telling anyone. we're building agentrail to make that journey visible and connect the dots. happy to chat if you want to dig into it.

[Harsh Raj @h_v_raj]

@JE4NVRG shipping the mcp bridge with the inbox on day one is the right call. can the agent see one customer's whole history when they've written on both whatsapp and telegram?

[Harsh Raj @h_v_raj]

@sonicaghi putting agent credentials and ai cost in the same place is the right call. the first question after a weird bill is which agent spent it. can konnect split spend by agent credential?

[Harsh Raj @h_v_raj]

crazy man!

matching openai's best embedding model without a single embedding means nothing to re-embed when the model changes.

[dale @daleverett]

Search is solved.

Life spent 4 billion years on it. Computers spent 70. Today, we've built an entirely new architecture of search that matches OpenAI’s best embedding model without using a single embedding.

Crack search, and you crack everything. So we have something even crazier coming next.

@wondersearchai is live. Backed by YC & Founder's inc

[Harsh Raj @h_v_raj]

the embarcadero

[Harsh Raj @h_v_raj]

9% of the companies on our agent readiness leaderboard grade an A.

the most common grade is a C. half of them sit there.

[Harsh Raj @h_v_raj]

product simulations so agents can try before they buy is where i'd start. today the trial is a sign-up form and a test key, and most agents need a human for both

[Mathew Pregasen @mathewpregasen]

I've recently made the claim that growth is becoming a first-order engineering problem. What does that actually mean?

Marketing used to be a matter of finding the right words, the right channel, and the right way to get noticed. You could polish a landing page, publish content, farm backlinks, and put up case studies to improve your growth.

Much of that still works today. But these practices were built for a world where a person did the research and made the decision.

Now, we're delegating decisions to machines. And machines don't evaluate products the way a person does. A good headline or a beautiful case study may get a person’s attention. But an agent doesn’t care about aesthetics. It's job is to be truth-seeking and make good decisions.

This is contrary to a lot of the movement in the market. Out of SEO, came GEO. Out of page navigation, came llms.txt. Out of backlinks came third-party mentions. But these are stepping stones. Do we really believe that the future of super intelligence is rewarding companies that throw the most articles into the ether? How rudimentary of a world that would be.

I believe that the future of growth hinges on truth and honesty. Machines are diligent. They’ll reward products that share evidence that helps them make better decisions. But that problem space looks radically different than today’s solutions.

In particular, that problem space becomes an engineering space. We will need systems that cryptographically cross-verify our usage, engagement, and customer loyalty without revealing exact numbers. We will need product simulations so that agents can try platforms before they buy. And we will need benchmarks that can be independently reproduced.

These new growth challenges are no longer marketing problems: they are engineering problems. This requires new tooling that looks completely different from what we have today. And it’s not merely marketing tools with a strapped on MCP. We need engineering tools—ones that speak the language of git, staging links, and code reviews—intently designed for growth engineering.

This materializes as a clear outcome. We need a new cohort of engineering tools spanning agents, harnesses, and IDEs for the next era of growth: one defined by building.

[Harsh Raj @h_v_raj]

@testingcatalog openai buying pierre for git workflows makes sense. coding agents can write the code, but the decisions behind a change are still scattered across issues, chats and reviews. that's the harder context to give codex.

[Harsh Raj @h_v_raj]

@dexhorthy congrats dex. agents running in the background with a team watching is where this was heading. good home for humanlayer

[Harsh Raj @h_v_raj]

@charlieholtz @natfriedman @alexandr_wang congrats charlie. running a few coding agents side by side is how a lot of us work now, and conductor made that normal before anyone else did

[Harsh Raj @h_v_raj]

three acquisitions today, none of them a model company.
openai is buying git hosting built for machines. conductor, which runs coding agents side by side, is joining meta. humanlayer's agent ide is joining factory.

the money is going to everything around the agent.

[Harsh Raj @h_v_raj]

@binsquares a coding agent gets the same attack without the fake interview. it clones whatever repo it's pointed at and runs the install script. network off by default fits there too

[Harsh Raj @h_v_raj]

@beknabdik @manufact 97 out of 100 across 1,114 servers is a real flex. which check cost speko the last 3 points?

[Harsh Raj @h_v_raj]

@gitghxst fewer retries is the one you can count. log how many tries an agent needs to get a call right and the friction shows up as a number

[Harsh Raj @h_v_raj]

sign up is where most tools fail this test today. the api works fine, then there's an email to confirm or a captcha, and the agent has to go find a human

[Mohak Bajaj @MohakBajaj5]

if your user has to go to a dashboard to use your product, he won't. he'd rather tell claude to rebuild it.

the fix isn't killing the dashboard. it's making everything in it reachable without one: a cli, an api, an mcp server, docs an agent can read in one fetch.

the test for any dev tool now: can an agent sign up, configure it, and read the error when it breaks, without anyone opening a browser tab. tools that fail it won't lose to a competitor. they'll lose to a weekend rebuild.