Kunal

There is no greater privilege than to be able to willingly push your mind and body to their limits every day

@shahahmed @gregpr07 @datadoghq Hey could we get beta access too?

@gregpr07 @datadoghq we do have one! available as preview to customers and GA soon. can I turn on for your company

what, @datadoghq doesn't have an MCP server?! I just want to run my background agents all the time exploring what goes wrong my infra.. how can I monitor my infra with agents?

@ntkris Single turn vs multi turn, the agent as an eval provides potential for a more robust eval or the optionality to construct metrics around the structure of the agent

@kunal_rye What’s the difference?

Over the next months I think we’ll see a trend from LLM-as-a-judge to agent as a judge or agent as an eval

Recently became aware of this pattern with Browserbase, they allow you to run an agent for a task the first time and the generate workflow code for subsequent runs. The challenge is that then you don’t have the adaptability on follow on runs that you do the first time, though I guess you could bake in a fallback yourself

I wrote about this a while ago here and im still not sure if it’s fully the way to think about it but I feel like it’s more right than wrong vtrivedy.com/posts/modern-p…

I may or may not be willing to die on this hill and maybe not even a hot take, but… 1. A LOT of the time, Agents should just be used as discovery mechanisms for workflows. True ASI is a perfect workflow generator for any Task 2. Like once my agent or my brain discovers a good workflow for my Task (ie. just a good series of steps that roughly works), the thing i almost always want is that straight up that workflow sprinkled in with ✨agenticness✨ to handle ambiguity (ie. the node may be an agent). Because then I get more reliability and I like that, and you prob do too 3. The vast majority of problems can be cast as workflows, but the UX of agents is way better. I just prompt the workflow in natural language like “do this, then do that, and then check this and go back to that…” 4. A lot of improvements in agent perf/reliability is basically just “workflowification” of agents enforcing upon them what steps to do and checking on that they do certain things in order. You may also call this “bringing your 🤌taste🤌/knowledge” to the problem basically this boils down to the fact that there’s a massive amount of useful economic work that isn’t open ended, it’s “roughly” some steps we need to follow so we should model the problem like that. Agents are THE way to handle truly open ended problems, but even then if you have some priors on that problem, workflow them in “But wait all you do is yap about agents?” 🤬 You right… Agents are fantastic “bridge nodes” at doing small, well scoped tasks. I think we all agree that agents rock at defined small tasks and are getting even better at medium tasks with no hand holding. This is coming from someone who sends Opus4.5 all day because the UX of agents is so good and that matters, not hating just saying what works well more often than not Agents are also fantastic at helping you discover the workflow, like sometimes you don’t even know what works well exactly so let them try stuff and then here’s the crazy thing…store that data, analyze it, and find out what works and make that more of a workflow cool so rant over, mostly an observation from seeing how reliability in agents often feels like workflowification and what that means for all of us building them

BREAKING: our AI Agents can now deep-research 10,000 accounts faster (and cheaper) than an Uber from JFK. Last month our growth team pushed this hard: 40,000 accounts deep-researched for product launches → personalized outreach → 15+ meetings and a closed-won in 30 days. Most GTM teams don’t lack ideas. They struggle with the cost and complexity of running them at scale. @kunal_rye , who leads our agent infrastructure, saw this firsthand. Adoption climbed past 1M runs per month, but cost was still the ceiling on pushing always-on workflows. After months of evaluating models against real customer workflows, Kunal & team shipped a major upgrade: Optimized agents with peak performance at a 10x cost reduction. Always-on intelligence in outbound becomes practical, not theoretical. What this unlocks: If you’re a cybersecurity company, Unify’s optimized agents can: (1) Monitor your entire TAM for breaches (2) Auto-qualify accounts if they’re B2B + using Auth0 (3) Use real breach context to write outbound that’s timely, relevant, human No manual research. No painful copywriting. No cost tradeoffs. The future of GTM is deploying 100,000 deep-research agents daily to understand every shift in your market. That future is closer than people think.

Despite some of the popular fears that all AI agent use-cases could get sucked into a single platform, I'd argue the other side of this for the enterprise. AI agents, possibly more than any prior era of tech, need to have a relatively high degree of specialization per domain or vertical. The model can be the same across fields, but the manifestation needs to be highly tuned. The reason is tied to what the AI agent is doing for the customer. What the customer is doing is "renting work" from the AI agent provider. This is similar to when a company either hires someone for a job or hires agencies or firms to help them.   When you hire people, you hire experts.  And when you hire consultancies or professional services firms, you hire a bunch of experts in a particular field.  There's a reason you tend to not hire people that are "just generalists", and why professional services firms tend to be optimized around focus areas, like tax, IT, legal, marketing, and so on. The consulting firm does everything either doesn’t exist or eventually specializes by practice area. The same is true for AI agents.  Companies are looking to solve problems in their workflow and business processes, and they're going to want experts to solve those problems, not generalists. You're no longer providing the tool for a person to do their work better, but you're actually supplying a worker to them. For anything important and value-added for that customer, they’re going to want the best agents that they can afford, similar to hiring talent in the rest of the market. Of course for lots of general purpose work this may not be the case, but for anything where their business is on the line it is. This dramatically increases the need for a deep domain understanding for the use-cases you're going after; custom UI that is tailored to the domain; access to relevant data just for the domain; and so on. The more general you are the worse off the results will be.  Of course there are nuances to this.  Generalists can do specialization if they divide things up to approximate specialization well enough. And equally, specialists can accidentally remain too small and not bite off enough of the problem for the customer. But either way, it's clear that specialization is going to win out in AI for the same reason it has in people.

when model performance stabilizes (we are in one of those periods right now) people start to evaluate for different objectives: cost, latency, customizing/differentiating, retaining IP

Announcing @unifygtm for PLG companies Product-led companies like @Perplexity and @Cursor are turning to Unify to convert their product usage data into revenue Most PLG founders wait too long to think about enterprise revenue. You scale users, hit PMF, then scramble to hire a revenue team when inbound demand stops supporting hockeystick growth 👇

Had the chance to do early testing of GPT-5, here's what we think!

Perplexity used @unifygtm to book 80+ meetings in 3 months without a single BDR we'd love to help your revenue team do the same s/o to @perplexity_ai and @jennysvng for being incredible partners

IN NEWS: @unifygtm has raised a $40M Series B led by @BatteryVentures. "In early '24, even as a seller, the best thing was to adopt AI to automate tasks." - @austinh___ "Our observation model gives memory to AI, creating a flexible context for better decisions." "It basically looks like you have a memory associated with every person you have ever talked to."

This morning I received a message from the most famous investor from Silicon Valley... Mr. Tres Commas himself. This video was produced for me by our friends at Unify who just announced their Series B fundraising. Congrats to you guys on the Series B and this super creative marketing. We're going to have to do something like this ourselves :)

Stoked to share this news, we have an incredible team here and we are hiring across all roles!

Update on @0xSplits Compliant Payments: We've now integrated @withpersona in our new recipient KYC flow. This allows you to collect W-8/W-9 forms from your recipients securely, compliantly, and accurately.

🚨 Introducing: Unify for Sales Reps ⚡ We’re bringing the power of AI automation directly to the frontline sellers who turn pipeline into revenue. Sellers today are forced to juggle tools, chase down context, and operate in silos. No more. Unify for Sales Reps is a new suite of AI-powered features designed to help reps move fast, stay focused, and book more meetings — with less effort. Today, we're excited to launch: 🧠 AI Research Assistant – Account research done for you 🧩 LinkedIn-to-Lead Extension – One click to enrich and sequence 📞 Manual + automated step sequences – Email, Call, LinkedIn, Action 📥 New Tasks Dashboard – One place for reps to execute, and managers to track All powered by Unify’s system-of-action for outbound sellers. 🎯 Let Unify can handle the busywork, so that your sales reps can crush their pipeline goals: unifygtm.com/blog/introduci…

Today we’re launching @unifygtm for Sales Reps. We don’t believe AI is going to replace people in GTM, we believe its going to make them 100x more effective 🧵

Unify's Connor Heggie and Kunal Rai share their learnings from building and running generalized research agents at scale. Watch the full session here: youtu.be/pKk-LfhujwI/?u… Catch up on all the talks from Interrupt: interrupt.langchain.com/video/?utm_med…

Learn more about how we use OpenAI's models to make growth a science here: openai.com/index/unify/

To continue on our mission towards enabling a way to engineer growth, we are excited to introduce our Observation Model, powered by @OpenAI o3 reasoning model, which underpins our multi agent system that constantly searches and surfaces signals and information about your potential prospects, helping your team take more informed actions.

At @unifygtm , we believe that go-to-market is a search problem. To help the best products win, companies need to find the best prospects for their products to succeed which requires sifting through large quantities of information and signals to determine the best fit.