Replicas

A mic drop announcement from me, we’re now offering $100 in @tryreplicas credit to EVERY YC startup in exchange for equity at a 50k capped MFN

Update for May 15th–19th, 2026 Major updates: - Google Workspace integration! Agents can now create/edit Docs, Sheets, Forms, and Drive files - Slack + Linear tasks now auto-select environments from context instead of always showing the picker - New “Unnecessary Comments Check” automation template for catching redundant or stale comments - Diff views now default to unified, with a split/unified toggle - Resizable workspace + file/diff sidebars with persisted widths - Workspace state popover now shows provisioned MCPs alongside skills, env vars, and files - Added The Context Company MCP template with prefilled Bearer auth formatting Misc: - Added public responsible disclosure policy + discretionary bug bounty program - Fix context usage pill displaying over 100% - Fix Global environment edits not invalidating pre-warmed workspaces org-wide - Added syntax highlighting to provenance inline file previews - Clarified Changes pane behavior for non-git-tracked files - Fix workspace “ready” toasts appearing before provisioning completed - Fix workspace provenance sidebar inconsistencies in multi-repo workspaces - Fix unrecoverable sandbox failures leaving workspaces stuck in Active state - Added workspace Error indicators + hidden chat composer for errored workspaces - Improved PR popovers with CI status, authors, comments, and diff stats - Fix ghost “+0 -0” diff card appearing above Claude Write/Edit tool calls - Fix org invite signup routing issues - Fix Global environment selection failures in Slack + Linear

Since doing this we had 4 of our all time highest shipping days by commits The velocity from background agents is unparalleled

im gonna have to review so many PRs man

what a self-maintaining codebase looks like: → User files a bug or feature request → Agent checks logs, audits the idea, opens a PR → Another agent reviews → Merge humans not in the loop at all. I think we're closer to this future than most people realize.

favourite thing about @tryreplicas is what it does to the team we pick up ways to be better engineers from watching how teammates prompt the agent, and everyone drops context + design notes on each other's PRs feels a lot like how tobi described river

David is right! However, if you’re not Sentry sized, you shouldn’t be building or hosting one in-house. There are many background agents out there that you can heavily customize to your company while still being generalized

Update for May 14th, 2026 Major updates: - Added Opus 4.7 1M - Full dashboard revamp, with a pane that shows all the changes and a file tree - Replicas can edit your Automations and Environments Misc: - Automations can now trigger on PR mergesages Automations can now trigger on PR merges - Automations no longer auto-disable after failures - Added filters/search to Automations - Added “Add to my workspaces” for shared fullscreen workspaces - Added “Add to my workspaces” for shared fullscreen workspaces - GitHub org connect now auto-connects personal accounts - Fixed disconnected agents being selected by default - Fixed duplicate workspace cards during creation - Fixed prompts disappearing while workspaces spin up - Fixed deleting workspaces during creation - Fixed shared workspace URLs hiding sidebar/chat UI - Fixed sidebar being capped at 100 workspaces - Fixed teammate messages showing as “You” - Fixed Linear bulk-assign retries on rate limits - Fixed Claude reasoning blocks rendering - Fixed onboarding losing CLI auth progress - Engine now auto-restarts after crashes

experimenting with getting replicas to function as part of the team - proactively taking stuff on if its in scope its going well so far...

we're switching to codex as the default @tryreplicas anthropic continually makes it harder for people to do cool things using your models and harnesses. it WAS a symbiotic relationship - people who loved Claude became so much more powerful with products like Replicas or OpenClaw or whatever can never understand why they limit it so much? its not like you aren't getting the training data you want out of the SDK...

THERE'S TWO OF US NOW I met Saai 3 years ago at the University of Waterloo, both in the double degree program. I even did my first startup with him. On the first day of @ycombinator, I was told that a cofounder is one of the most influential people in your life. They are the person who will be there for the highest highs and the lowest lows of the company. I didn't take it seriously (because I didn't have a cofounder). But I understand now. Since the minute @SaaiArora has joined, I haven't felt anything other than excitement at how much higher our ceiling has become. The dream is twice as ambitious, and the probability we make it a certainty. I've been pitching for Saai to join @tryreplicas nearly since the day I created it. Now he's here, and we're about to go 10x faster ;)

@SaaiArora 👀

A pretty unexpected life update: A few days before my Ramp Applied AI internship was supposed to start, I made the difficult decision to step away and join @tryreplicas as CTO in the current YC P26 batch. This was honestly one of the hardest career decisions I’ve made. Ramp is one of my favorite companies in tech, and every interaction I had with the team reinforced that. The people, culture, and engineering quality there are genuinely exceptional. I also want to emphasize the immense respect I have for the Applied AI team and the work they’re doing around Glass, Inspect, and agentic systems more broadly. Watching companies like Ramp push forward on background agents and self-maintaining software has completely changed how I think about the future of engineering. What ultimately solidified the decision was realizing that @connortbot and I share the same vision for where software is heading. We’ve worked together before, and joining him as co-founder felt like a natural fit. I believe the future of software is persistent background agents that can reason, investigate, write code, debug systems, and continuously improve autonomously, dramatically amplifying what small teams can build. Excited for what’s next :)

Introducing Custom Environments - unlimited use cases for background agents. With each custom environment, you can configure it with system prompts, skills, MCPs, and dependencies/tooling tailored for the job. For example: - a Debugging environment that has access to Cloudwatch logs and our traces - a Data Analyst that has Postgres & Snowflake - a Marketing environment instructed to read X and LinkedIn posts - a ML environment with access to an H100 and datasets - a Sales environment with access to HubSpot or Salesforce Our customers are already automating engineering - E2E tests with specific prompts and access to latency metrics, adjusting classifiers, With 3 sentences and a click, Replicas sent this to me on Slack:

The best thing about @tryreplicas is that it plays hardball to make sure the PR stays green. - Responding to code review comments - Restarts flaky unit tests - And it documents all the custom fixes it does

replicas just made an account on replicas and did the 2fa on its own several launches incoming 👀

its been like 2 days and hes not even sending me the PR links himself anymore we're all just replicas proxies

@SaaiArora @sarahfim @conductor_build We would also like to know 👀

@sarahfim @conductor_build @tryreplicas Fair enough, did you end up trying @tryreplicas

what is the best, thin, agent manager for claude code & codex? i think t3Code is my favourite so far, but i have not tried that many

Replicas just built an audit logs feature, ran our codebase locally and tested it out then sent us videos & screenshots in our slack channel we are 🤏 this close to automating out QA too