In partnership with
Three repos launched this week, same idea from three different angles: stop treating your agent like a chat window and give it a place to work. OpenBot hands every agent a computer of its own. Cumora gives agents a seat in the team chat. macOS Harness hands one the whole Mac. A month ago the argument was which model writes better code. This week it moved to where your agents actually work, and that is a much bigger question.
The AIgent stays free for everyone, and reader support is what keeps it that way. Support The AIgent so we can keep delivering it to you. Your support funds the newsletter and the people and technology that make it.
|
Sponsored
|
You're already late hiring a marketing agency for 2027.
Your 2027 budget is almost finalized, and the clock is ticking to hire your new agency partner for next year.
An agency search takes this long:
3 weeks to scope the brief and build a shortlist
4 weeks of pitches and reference calls
3 weeks for contracting and legal
60–90 days of onboarding before the work is worth judging
That's over 5 months wasted, so an agency hired in January doesn't deliver results until May.
Vendry helps you hire the best agency for your business, ASAP.
Describe the brief, we do the shortlisting, and your search begins tomorrow.
Tell us what you need and we'll send 3 or 4 vetted agencies matched to your budget and category in <7 days.
The Drops
RepoOpenBot
2,501 stars · CopilotKit/OpenBot
AI coworkers you can hand real work to. Each one gets a computer of its own: a real browser with its own logins, its own files, and only the tools you grant.
Here is the real problem it solves. Agent access today is all or nothing: either the bot lives in a chat box, or you hand it your whole machine and hope. OpenBot decides every action before it happens and records it after, so you can give an agent real accounts without giving it everything.
Quick start: bring any AG-UI agent, or start with theirs; it shows up as a coworker with a computer you can watch.
Catch: six days old and alpha. Pilot it on scoped accounts, not your ops backbone.
AffiliateReclaim.ai
An agent that can take on real work is only useful if you actually have deep-work hours left to hand it a task, which is its own scheduling problem. Reclaim.ai, an AI calendar that auto-schedules your tasks, habits, meetings, and focus time into Google or Outlook, then reshuffles everything around conflicts as they land. If you build and context-switch all day, it defends your deep-work blocks so you do not have to.
Use it for: defending your deep-work blocks while the calendar reshuffles itself.
We may earn a commission.
Repocumora
2,953 stars · yetone/cumora
Team chat where AI agents are first-class teammates: same roster, same DMs, same group chats, same Kanban board as the humans.
Use it for: running a mixed human and agent team in one room. Agents hold personas and memory, claim work, coordinate without colliding, and send real email.
Catch: teammates this autonomous need scoped access. Start them on the cloud pods or a spare machine, not your main accounts.
Repomacos-harness
728 stars · browser-use/macos-harness
The thinnest harness that gives an LLM complete freedom on a Mac: one Python process wired to macOS, your real browser, and your files.
Use it for: automating the Mac apps nobody built an integration for. When a helper is missing, the agent writes it mid-task in plain Python.
Catch: complete freedom cuts both ways. Run it on a Mac you can afford to let an agent drive.
MCPTencentDB Agent Memory
23,992 stars · TencentCloud/TencentDB-Agent-Memory
A team-level memory hub that turns agent conversations, docs, and code into four reusable memory assets.
Real problem: most agent memory dies with the session. This gives a team a shared, queryable memory layer instead of everyone's context living in their own chat history.
Quick start: point it at one agent's transcript log and see what it extracts before wiring it team-wide.
Catch: it's new. Treat the memory it builds as a draft, not ground truth, until you've audited a week of it.
|
From Our Partners
|
Bring OOH Into the Modern Marketing Stack
AdQuick makes Out Of Home advertising approachable, measurable, and performance-focused. Designed for marketers at startups and large brands alike, it combines digital efficiency with real-world reach—so your campaigns always hit the mark.
Start Here
Paste a sample of "good" instead of describing good.
Here's the trap most people fall into when they ask an AI for help: they try to explain what they want in words. "Make it sound professional." "Write something punchy." The AI guesses, and the guess is usually generic, because "professional" and "punchy" mean something different in your head than they do in the model's training data. There's a faster way.
1. Find one example of the thing you actually want, an email you liked, a paragraph with the tone you're after, a spreadsheet formatted the way you need it.
2. Paste that example into the chat first, before you ask for anything.
3. Say "match this style" or "use this format" instead of describing the style in adjectives.
4. Then give your actual request.
5. If the output drifts, paste a second example and say "more like this one, less like the last one."
TRY THIS: Open your last chat with an AI tool. Find one sentence you've written elsewhere that has the exact tone you wanted. Paste it in and ask the AI to write your next thing "in this style." Compare it to what you'd have gotten from just describing the tone.
|
Recommended
Once you've got one paste-ready sample of good, the fastest next move is borrowing more good samples instead of writing your own from scratch. HubSpot's 100+ ChatGPT Prompts guide is a free download of working prompts for writing, research, and daily tasks, organized so you can lift what you need and go. We may earn a commission. |
Frontier Signals
Linus Torvalds ran a debug session with an AI doing the grunt work, and it told him flat out a fix was impossible. For operators, that's the tell worth noticing: the value wasn't the AI being right, it was the AI being honest about the limits of its own guess. (Simon Willison)
Vercel built a benchmark that runs every major coding agent against the same infrastructure instead of each harness's own sandbox. If you're choosing between Codex, Copilot, or a dozen others, this is the first apples-to-apples comparison worth trusting. (Vercel)
Simulation is replacing real-world training runs at a fraction of the cost, roughly 10% worse results for 100x cheaper and thousands of times faster. For an operator watching compute bills, that tradeoff is the one to know about before your next fine-tune. (Latent Space)
LinkedIn's "Seems like AI slop" button has been clicked over a million times since launch. The platform built a public shame mechanism for the exact content style every AI writing tool defaults to, worth remembering before you auto-post anything. (The Verge)
Waymo is now outspending Uber on lobbying to clear a federal path for fully autonomous taxis. The regulatory fight, not the tech, is what decides which robotaxi fleet you'll actually be able to hail first. (Ars Technica)
|
Recommended reading
If you like The AIgent, a small group of operator-tier publications worth your inbox: see the shortlist. |
Which best describes you right now?
Before You Go
Reply and tell me which of these you would put to work first: a coworker with its own computer, an agent in your team chat, or one driving a Mac. I read every one, and the best replies end up as the next issue's opening line.
See you Tuesday.
Before you go: we started a room for people actually building with agents. This week's "I'm stuck" thread is pinned in General: post the thing blocking you and we answer every one. Join the community →
|
Want to reach builders shipping with AI every weekday? Advertise in The AIgent. |
