Signal #61: OpenAI and Anthropic work the night shift

Good Friday, friends.

Last week the story was one prompt doing a whole job while you watched. This week the job runs while you are not in the room. OpenAI and Anthropic both shipped agents that work on a schedule you set once, across the apps you already live in, and hand you the output instead of asking for the next instruction. The skill that matters is no longer how fast you prompt. It is how well you write the standing brief and how carefully you check what comes back.

If you run any recurring workflow, the ChatGPT Work and Cowork releases are the reset for how you spend your mornings, and the Playbook below is how to stand up your first one. If you pick tools for a team, the Superhuman Docs relaunch shows where the workspace category is heading. If you build or invest in CEE, the Oxylabs round is proof you can reach €3.1B on cash flow. And if you buy compute, Nvidia's new financing model is the clearest sign yet that the boom is being underwritten in a circle.

On July 9 OpenAI launched ChatGPT Work, an agent that pulls from your connected apps (Slack, Teams, Google Drive, all @-mentioned) and keeps working for hours after you close the laptop. Two days earlier Anthropic brought Claude Cowork to web and mobile, where scheduled tasks now run with no device online at all. Set Monday's client prep for 6am and Claude works through the email threads, transcripts, and recent news, then leaves the briefing built and the follow-up drafted but unsent.

Two of the three biggest labs shipped the same primitive inside 48 hours: an agent that clocks in on a cadence, not on your keystroke. For two years the pitch was speed, a tool to make you quicker at what you already did. The night-shift model breaks that. The work now happens on a schedule, inside the systems you already use, and you meet the result rather than drive it.

Your edge moves accordingly. It stops being how fast you type and becomes how clearly you define the outcome and how well you check the result. Pick one task you now do by reflex, the Monday pipeline review or the weekly inbox triage, and hand it to a scheduled agent this week. Judge it on the ugliest version of the job, since that is the run that decides whether you can stop doing it yourself.

Still using GoDaddy or Namecheap? There’s a better way with Porkbun!

Porkbun is the domain registrar trusted by creators, developers, entrepreneurs, and folks who want low prices without the nonsense.

Why people are choosing Porkbun:
• Most domains sold at cost
• Low, transparent registration and renewal pricing
• Free features like WHOIS privacy and SSL certificates
• Powerful web and email hosting options
• Real human support 24/7, 365 days a year
• Named the #1 domain registrar by Forbes Advisor and USA Today

For launching a business, building a personal brand, starting a side project, or creating your first website, Porkbun makes it easy.

Get $1 off your next domain registration with Porkbun now.

Get Your Domain Name Now

The gap between "an agent could do this" and "an agent does this every Monday" is one careful afternoon. Here is the order that works.

Pick a task with a clear done. Do not start with "run my inbox." Start with one bounded job that has an obvious success and an obvious failure: the weekly revenue pull, a Monday account brief, a standing competitor scan. A narrow scope is what lets you write instructions the agent can follow and what lets you tell whether it worked.

Write the standing brief, not a prompt. Describe the outcome, the inputs, and what "good" looks like, then save it as a repeatable task rather than retyping it each week. Keep what the agent does separate from what it reads, so you can update the source without rewriting the job.

Connect only the apps the job needs. Both tools act across your stack, so wire in the two or three sources the task touches (the CRM, the shared drive, the calendar) and nothing else. Fewer connections means fewer ways for the run to wander.

Leave the output drafted, not sent. For anything that leaves the building, tell the agent to stop one step short: draft the email, build the doc, stage the report, but wait for your approval. Cowork's 6am client-prep example does exactly this, and it is the cheapest safety rail you have while you learn to trust the run.

Read the first three outputs before you rely on it. Run it on schedule, then check the early results against the real thing. Watch where it stalled or reached for the wrong source, fix that one step, and let it keep going.

The move this week: take one recurring task off your own plate, write down what "done well" means for it, and schedule it. If the first three runs clear your bar, you have found your first job to hand off for good.

On July 8 Superhuman relaunched Coda as Superhuman Docs, folding the doc-and-table tool into its productivity suite behind a chat assistant that acts across the whole document. Docs AI drafts content, builds and updates tables, organises data, and resolves comments from a single panel. Every existing Coda doc converts automatically, pricing holds, and an MCP connection wires ChatGPT, Claude, and Cursor into your docs with one click.

The consolidation is the tell. After the Grammarly-era roll-up, the workspace category is collapsing into a few AI-native suites, and the document is being rebuilt so an agent can operate inside it rather than autocomplete on top of it. For a team choosing a knowledge tool, that one-click MCP hook matters more than the writing features: it means the same scheduled agents from the night-shift story can read and edit your docs directly, not through a copy-paste.

If you already run Coda, you have this today, so point Docs AI at one document you rebuild every week and let it own the tables. If you do not, it is now a cleaner test of the "agent lives in the doc" pitch than the alternatives. Superhuman Databases, up to a million rows per database, is in private beta if you outgrow the page.

On July 9 Vilnius-based Oxylabs took its first outside investment in eleven years, €113.6M ($130M) from Warburg Pincus at a €3.1B ($3.6B) valuation. The company has run profitably since 2015 on web-data infrastructure, the pipes that let software pull live data off the open web, and now posts roughly $350M in annual recurring revenue. It is Tesonet's second bootstrapped unicorn out of Lithuania after NordVPN.

The AI hook is direct. Oxylabs sells itself as the data layer for the AI era, the live-web fuel that agents need to do anything useful, and every scheduled agent that checks a price, reads a competitor, or pulls a filing runs on infrastructure like this. The regional lesson is louder than the round. You can build to $3.6B on cash flow, from the Baltics, and let the fund come to you on your terms rather than raising to survive.

If you are building in CEE, the Oxylabs and NordVPN pattern is a real alternative to the dilution-first path: stay profitable, own a boring category, take growth capital only when it strengthens your hand. If you build agents, note where the AI-era margin is quietly sitting, in the unglamorous plumbing that everything above it depends on.

On July 1 Nvidia unveiled a revenue-share financing model for the smaller clouds that buy its chips. It backstops their GPU deployments, renting back unused capacity at a fixed rate in exchange for a cut of the cloud revenue those chips generate. The first takers, Sharon AI and Firmus, committed 210,000 Grace Blackwell GPUs. Markets read the circularity fast: Sharon AI's stock fell about 14% the next session.

Nvidia now sells the hardware, de-risks the loan behind it, and earns a slice of the usage on top. That is the same "assign the outcome and take a cut" logic as the agent products, applied to capital: the vendor climbs the stack and earns a recurring line where it once made a one-time sale. Expect other infrastructure sellers to copy the shape, and expect harder questions about how much AI demand Nvidia is now financing itself.

If you run a smaller cloud or buy compute, the math just changed: cheaper access now, a revenue share forever, and deeper dependency on the firm that sets the price. For everyone else, it is the clearest sign the compute buildout is being underwritten in a loop. Watch whether the next neocloud raise leans on a vendor backstop instead of a bank.

Five tools to install or test this week. Productivity carries two, since it is the theme of the whole issue.

Productivity: OpenAI's voice now listens and talks at once. ChatGPT Voice moved to GPT-Live on July 8, a model that listens and speaks simultaneously, so interruptions and turn-taking feel natural instead of walkie-talkie. It turns hands-free dictation and walking-meeting capture into something you can lean on. Try it for your next drive: talk through a rough draft and let it push back in real time.

Productivity: Claude will now show you your own AI habits. Anthropic launched Reflect on July 9, a beta dashboard that visualises what you use Claude for, when, and how much, with nudges toward what you might want to keep doing yourself. It is a sharp gut-check as delegation ramps up. Turn it on and decide, deliberately, which one task you will not hand off.

Design: Gamma steps out of decks into brand assets. Gamma launched Imagine on July 6, generating on-brand logos, infographics, diagrams, and social posts as standalone assets, not just slides. It is a fast marketing-ops win without opening a separate design tool. Spin ten variations of one asset you rebuild every week and keep the best.

Dev: GitHub Copilot gets a model swap and an MCP store in JetBrains. Copilot in JetBrains IDEs can now run OpenAI's Codex as its agent and browse, add, and start MCP servers from an in-editor marketplace instead of hand-editing JSON. Update the plugin, wire up one MCP server, and try the Codex agent on a low-risk refactor.

Dev: Grok 4.5 undercuts the frontier on coding cost. xAI shipped Grok 4.5 on July 8, a coding-and-agent model that ranks top-four on the Artificial Analysis index at roughly 60% less per task than Opus 4.8 or GPT-5.5, and it is live in Cursor now. If you run coding agents at volume, point one real, messy task at it and price the result against your current default.

Where the region and the wider market gather this autumn. Book the ones worth the flight before the good hotels go.

Startup Fair, Sept 17, Vilnius. The Baltics' largest startup event, with a pitch battle putting 40+ startups in front of 200+ investors.

Bits & Pretzels, Sept 28-30, Munich. Europe's founders' festival, roughly 7,500 attendees including 1,500+ VCs, capped with an Oktoberfest finale.

How to Web, Oct 6-8, Bucharest. Eastern Europe's biggest startup conference, drawing leaders from OpenAI, Meta, and Google plus 250+ investors. The most CEE-relevant stop on the calendar.

World Summit AI, Oct 7-8, Amsterdam. The 10th edition, Europe's premier AI-specific conference, anchoring World AI Week.

Wolves Summit, Oct 14-15, Kraków. Tight CEE corporate dealflow, roughly 100 investors and 250 founders, plus the European CVC awards.

Web Summit, Nov 9-12, Lisbon. The continent's largest tech gathering, 70,000+ attendees from 160+ countries.

Slush, Nov 18-19, Helsinki. The most investor-dense flagship in Europe, the one to attend if you are raising at Series A or deploying it.

Next edition soon,

Çelik