@_aj

Building @GrepdotAI (YC F26). Prev Product @compound, @brexhq, @coinbase, @google. 2x @ycombinator founder, @uniofoxford. Dad of twins, in-person maximalist.

San Francisco
Joined January 2008
Pinned Tweet
We're building self-improving agents to automate tedious, repetitive operational work @GrepdotAI. At IG, a UK online brokerage, Grep agents review compliance alerts, helping over 80% of traders onboard without manual intervention. IG is now a FTSE 100 company. Operational work often means following a specific procedure, gathering information from different sources, checking it against what's needed, and deciding what to do next. The steps are repetitive, but no job is the same, making the work tedious to do by hand and hard to automate with workflows alone. Meanwhile, using frontier agents to reason through every task is often too expensive to scale. Grep.Al gives you an agent workforce that: -> Takes minutes to create and are domain experts with the right skills and tools from day one. -> Learn from every run, moving repeatable steps into code and workflows and saving frontier intelligence for exceptions and unfamiliar cases. -> Get smarter and cheaper as they scale using our AgentRun harness. -> Runs on our managed agent platform with secure sandboxes, audit logs, evals, cost controls, team sharing, and access controls included. Grep AI agents are available on demand and can run on schedules or in batch. Access them via the console, MCP, API, or SDKs that integrate into your existing workflows. Our customers, including IG, Airwallex, Worldpay, FV Bank, and Citizens Bank, have saved more than 250,000 hours of operational work with Grep Al agents across onboarding, doc verification, compliance, underwriting, operations, HR, GT ops, market research, and business intelligence. If operational bottlenecks are slowing your business down, create an agent workforce with Grep to automate them in minutes. P.S. We're in the @ycombinator Fall 2026 batch!
47
28
9
177
76,466
Day 3 of open sourcing our harness as pi-extensions: pi-durable-disk and pi-transport. We use both to run agents from anywhere. For us, the agent is the workspace. And the workspace lives in @archildata. With @pidotdev durable, the session DB also lives in the disk. We can mount and unmount it from its sandbox, and run the pi loop from a web browser (w WASM) or a @modal VM with GPUs - sometimes in the same session. Even easier, with Archil's serverless execution, we can run simpler agents. For example, we have an agent in-process go through 100s of GBs in Claude Code sessions and recall the entire history of a project in seconds.
14
11
2
92
5,930
Day 2: One of our most common requests from customers was to automate internal tools and cloud software that didn't have APIs or MCPs. Last week we launched that feature in @grepdotai. Today we're open-sourcing the harness feature that makes it work as an @pidotdev extension.
Day two of open-sourcing agentrun as @pidotdev extensions: meet pi-browser. Give your agent a browser with persistent context. We use pi-browser plus pi-vision to do complex website analysis. Powered by @browserbase.
2
1
6
229
Honestly I don’t want to check out any part of a codebase. I want to just point my agents at a codebase, tell them to make changes, and let agentic checks and evals pass. If I’m running my agents in the cloud, and the code is hosted in the cloud, why check out anything at all?
1
82
Another week of launches! We're open-sourcing the best parts of our Jev-powered harness as Pi extensions, starting with pi-vision.
A few weeks ago, we built a new harness, AgentRun, which turns repetitive steps into workflows and code using Jev. This week we're open-sourcing its best features as @pidotdev extensions! Day 1: Pi-vision uses Jev to give your agents eyes without wasting frontier tokens. Analyzing videos, images, or documents (a common use case @grepdotai) with frontier models is extremely token-inefficient. With this extension, @typesafeai's Jev does the analysis, while your Pi agent does the orchestration. The result is a faster and cheaper workflow for detection and analysis at scale! You can use this workflow for document verification, understanding image content, and needle-in-a-haystack search in videos.
1
1
11
526
Prediction: every decade will bring a new, even more impressive Roadster demo, but the actual product will remain elusive. And maybe that's actually ok? It's a symbol of technological progress and our aspiration to build awesome stuff.
3
3
334
OK I have to say @HyperFrames_ Studio is really great. I went from being bearish on AI-powered motion design to doing 5 product marketing videos in 5 days with this tool. Hats off to @Rames_Jusso and the team! Here are all the videos I made with Hyperframes 🧵
6
2
1
18
1,071
Day 4: What's an agent without evals? Today we're launching native evals in @GREPdotAI. Now every time you create an agent, we automatically create a set of evaluation questions that you can edit and update. Then, while the agent is running, your team can evaluate it and provide feedback on its accuracy. That feedback goes back into the agent harness, and Grep suggests prompt updates you can apply like a GitHub pull request to improve your agent.
1
63
Day 5 of our week of launches: our new and improved MCP for @GrepdotAI. Everything you do with Grep agents now works from Claude, ChatGPT, Claude Code, Codex or any MCP client. → Build agents from a sentence: Grep designs the prompt, skills and tools, previews its plan, keeps every version and shares it with your team. → Run them from any chat: the agent's form opens in the conversation, you see the price before it starts, steer it while it works and collect every file it makes. → Stay in control: every run is capped, and your credits, usage and API keys are one question away.
63
Every day this week, we're launching a new feature @GrepdotAI to help you automate the work that's slowing down your business. 𝗗𝗮𝘆 𝟮: 𝗗𝗲𝗰𝗶𝘀𝗶𝗼𝗻 𝗼𝘂𝘁𝗽𝘂𝘁𝘀 𝗽𝗼𝘄𝗲𝗿𝗲𝗱 𝗯𝘆 𝗝𝗲𝘃 Customers often ask why our agents don't attach confidence levels to their recommendations. My answer has been that LLM confidence levels don't mean much. The model writes a number after the answer, predicting what a confident-sounding score looks like, with a bias toward making humans happy. RLHF pushes models overconfident. Decision models like Jev by @typesafeai change that. You declare the options up front. Jev returns a probability for each one in a single pass, plus a confidence score derived from how concentrated that distribution is. TypeSafe trained it with RLCD, which rewards calibration instead of rater approval, so decisions at 0.9 should be right about 90% of the time. In Grep, you can now build agents that use frontier models for the research and diligence, then hand their findings to Jev to make the final decision against your criteria. You set the threshold: above it the agent acts, below it a human reviews.
1
2
79
AJ Asver retweeted
Proud to power such an awesome product!
Day 3 of our week of launches: Connected accounts + authenticated browser access. Grep already includes over 100 data integrations out of the box - no logins needed. But as our agents took on more automation, it became clear that customers needed a way to add their own integrations, APIs, and even internal tools with no MCP or API connection. That's why we're excited to team up with @composio to bring over 1,000 integrations to @GREPdotAI for all your favorite SaaS products, including Gmail, Calendar, Outlook, Notion, Salesforce, and much more. Customers can also add their own custom APIs and MCPs to Grep's agents for tools without a built-in integration. But not every product or application has an API or MCP, like internal tools, legacy software, and government-run databases. For those use cases, we're also launching authenticated browser access for agents. Just log in to any web app once, and we securely store your credentials in an encrypted vault so your agents can use the application for automation. If you don't want your credentials stored, you can also log in manually every time the session expires. Now, Grep agents can automate any online application or website and access any online data source, reducing the repetitive work that slows down your business. For example, a customer onboarding onto a fintech might hit a manual review step because a compliance screening alert flags their name as similar to one on a sanctions list. An agent can pull compliance alerts from a case queue in Salesforce, log in to a legacy screening solution like Nice Actimize to disposition the alerts using deep research and Grep's data integrations, and then send RFIs to customers via email when there isn't sufficient evidence. Check it out at Grep.ai
3
5
18
2,343
Day 5 of our week of launches: our new and improved MCP for @GrepdotAI. Everything you do with Grep agents now works from Claude, ChatGPT, Claude Code, Codex or any MCP client. → Build agents from a sentence: Grep designs the prompt, skills and tools, previews its plan, keeps every version and shares it with your team. → Run them from any chat: the agent's form opens in the conversation, you see the price before it starts, steer it while it works and collect every file it makes. → Stay in control: every run is capped, and your credits, usage and API keys are one question away.
7
2
3
26
6,970
Connect in a minute: add api.grep.ai/mcp as a custom connector and sign in with your Grep account. No API keys to paste.
90
AJ Asver retweeted
Day 4: What's an agent without evals? Today we're launching native evals in @GREPdotAI. Now every time you create an agent, we automatically create a set of evaluation questions that you can edit and update. Then, while the agent is running, your team can evaluate it and provide feedback on its accuracy. That feedback goes back into the agent harness, and Grep suggests prompt updates you can apply like a GitHub pull request to improve your agent.
8
1
3
15
912
Day 4: What's an agent without evals? Today we're launching native evals in @GREPdotAI. Now every time you create an agent, we automatically create a set of evaluation questions that you can edit and update. Then, while the agent is running, your team can evaluate it and provide feedback on its accuracy. That feedback goes back into the agent harness, and Grep suggests prompt updates you can apply like a GitHub pull request to improve your agent.
8
1
3
15
912
Day 3 of our week of launches: Connected accounts + authenticated browser access. Grep already includes over 100 data integrations out of the box - no logins needed. But as our agents took on more automation, it became clear that customers needed a way to add their own integrations, APIs, and even internal tools with no MCP or API connection. That's why we're excited to team up with @composio to bring over 1,000 integrations to @GREPdotAI for all your favorite SaaS products, including Gmail, Calendar, Outlook, Notion, Salesforce, and much more. Customers can also add their own custom APIs and MCPs to Grep's agents for tools without a built-in integration. But not every product or application has an API or MCP, like internal tools, legacy software, and government-run databases. For those use cases, we're also launching authenticated browser access for agents. Just log in to any web app once, and we securely store your credentials in an encrypted vault so your agents can use the application for automation. If you don't want your credentials stored, you can also log in manually every time the session expires. Now, Grep agents can automate any online application or website and access any online data source, reducing the repetitive work that slows down your business. For example, a customer onboarding onto a fintech might hit a manual review step because a compliance screening alert flags their name as similar to one on a sanctions list. An agent can pull compliance alerts from a case queue in Salesforce, log in to a legacy screening solution like Nice Actimize to disposition the alerts using deep research and Grep's data integrations, and then send RFIs to customers via email when there isn't sufficient evidence. Check it out at Grep.ai
35
12
6
65
11,004
Our coding agents mention each other in Slack without being prompted. They ask for support / context (each agent has access to their owners' sessions and the company brain). It's a bit eerie, tbh.
2
1
5
456
Today we're launching artificial intuition. A model that predicts which idea will stand out, before anything is published. Already live in Native for 2,200+ businesses. API waitlist is open: native.no/norn 🚀
65
19
6
211
11,532