Four releases this week each folded an entire workflow into a single instruction: a spoken description becomes a live phone agent, trading agents run billions on their own, one cheap API turns a photo into video, and Microsoft ships its own coding model inside Copilot. The operator's job is moving from wiring tools together to naming the outcome and checking the result.

Morning, friends.

For two years the pitch on every AI tool was the same: it makes you faster at the thing you already do. This week the pitch changed. The tools stopped speeding up your work and started doing the work, if you can describe the outcome cleanly enough. That is a different skill, and it is the one worth practising now.

Below: xAI's builder that turns a paragraph into a working phone agent, a Prague lab that hands trading agents real money, Google's one-API path from photo to video, and Microsoft quietly shipping its own coding model. The through-line is that you are no longer buying tools. You are assigning jobs, and your edge is how well you brief and verify.

On July 1 xAI opened Voice Agent Builder in beta, a no-code way to stand up a production phone agent in about two minutes. You write a plain-language description of how calls should go, attach your documents and tools, and it handles telephony, retrieval, guardrails, and call recording out of the box. The pitch is against the standard voice stack, which stitches together three separate APIs for speech-to-text, reasoning, and text-to-speech, each one adding cost and a new way to fail.

The shift is that voice stopped being an integration project. Grok Voice is one model on a speech-to-speech path, and xAI claims it clears its own τ-voice benchmark at 67.3% against 43.8% for Gemini 3.1 Flash Live and 35.3% for GPT Realtime. At $0.05 a minute with voices included, a support or booking line that used to need a vendor and an engineer is now a paragraph and an afternoon.

If you run any phone workflow, front desk, order status, or appointment booking, build one this week and call it yourself. The tell is not the benchmark, it is whether the agent survives your ugliest real call. Give it your hardest workflow, not your cleanest, because that is the one that decides whether you can retire the human queue behind it.

Every manager faces the same situations before lunch: a message to land, a meeting to run, a hiring call, a report due. The AI Report built 37 Claude prompts for exactly those moments, organised by the situations every manager faces. 

Copy the prompt, fill the brackets, run it in Claude, and get back 10+ hours a week. Oh, and it's free. 

All you have to do is subscribe to The AI Report, a 5-minute daily AI brief read by 400,000+ business leaders at IBM, AWS and Microsoft, and the full prompt pack lands in your welcome email. The newsletter and the prompts, both free. Subscribe and grab both

Subscribe and get 37 prompts free

The gap between "I could build a voice agent" and "I have one answering calls" is now an afternoon, if you work in the right order. Here is the shape of it.

Pick a narrow call, not a broad one. Don't start with "handle all customer support." Start with one bounded job that has a clear success and a clear failure: order status, appointment booking, or after-hours triage. A narrow scope is what lets you write a description the model can follow, and what lets you judge whether it worked.

Write the flow as instructions, then load the knowledge. In the builder you describe how a call should go in plain language, then attach the documents the agent answers from (your policies, product specs, an FAQ) as a shared collection instead of pasting facts into the prompt. Keep what the agent does separate from what it knows, so you can update the FAQ without rewriting the agent.

Wire the actions before the voice. The agent only earns its keep if it can act, so connect the two or three tools the job needs first: a calendar for booking, an API for order lookups, a clean hand-off to a human when it hits its limit. Voice and phone number come last, and you can test the whole thing in the browser before it ever rings a phone.

Set guardrails, then break it on purpose. Add limits for what it must never do (read back card numbers, wander off script), then call it with your worst case: a bad connection, an interruption, a caller who changes their mind mid-sentence. Every call is recorded and transcribed, so watch which tool it reached for and where it stalled, and fix that one thing before the next test.

The move this week: take one call type you currently pay a person to sit on, build the agent for exactly that, and run ten real calls through it. If it clears eight, you have found your first job to hand off.

A Prague lab just took the "assign the outcome" idea to its logical end. EquiLibre Technologies, founded by three ex-DeepMind researchers, is now valued at more than $500 million after a Series A that Creandum called the largest single check it has ever written. The same team built DeepStack, the first AI to beat pros at no-limit poker, then aimed that reinforcement-learning approach at markets. Working with quant firm Tower Research, its agents have traded billions in daily volume across the S&P 500 and Nasdaq.

What makes this the story of the week is the scoring. CEO Martin Schmid puts it plainly: "the scoring is super simple: how much money did the agent make?" You do not brief these agents on tactics; you hand them an objective and a live market and let them learn. EquiLibre claims a perfect record of zero negative months since inception, built out of Prague with 25 people, not Mayfair or Menlo Park.

Note where the talent chose to sit. Schmid says it is easier to keep good people in Prague because "there's not a new sexy AI thing happening every two months," and the firm is now building one of the largest compute clusters in CEE. If you are hiring frontier talent, the opening is a serious problem in a place people don't churn out of. Watch whether more DeepMind and OpenAI alumni route their next company through the region instead of around it.

On June 30 Google put its cheapest image and video models in one API. Nano Banana 2 Lite renders a 1K image in about four seconds for $0.034, and Gemini Omni Flash generates and edits video from text, image, or video prompts at $0.10 a second. The point Google keeps making is that you chain them: generate a still, pass it straight into Omni Flash, and animate it, all inside one session.

The collapse here is the creative stack. A product shot that used to move through a generator, an editor, and a separate video tool is now one prompt to one vendor at drafting-budget prices. Cheap and fast changes what you make, not just how you make it. At three cents an image you stop rationing and start generating fifty versions of an ad, a listing, or a thumbnail, and picking the winner.

If you produce any volume of visual content, run one real job through the pair this week, a batch of product images animated into short clips, and price it against your current workflow. The limits are real: Omni Flash tops out at ten-second clips today, so treat it as a drafting and iteration engine, not a finishing one. The teams that win here will be the ones who rebuild their content pipeline around cheap iteration instead of guarding a slow, expensive one.

On June 26 Microsoft made its own coding model generally available inside GitHub Copilot Business and Enterprise. MAI-Code-1-Flash is built by Microsoft's own AI group, purpose-built for code, and tuned for the fast, high-volume loops that agentic coding runs on. For years Copilot ran on other labs' frontier models. Now Microsoft is routing its highest-volume surface through a model it owns.

The move worth reading is the boundary quietly dissolving between the tool and the model underneath it. When your IDE vendor also builds the model, "which model" stops being your decision and becomes theirs, folded into the product and the bill. That is efficient when it works and a lock-in risk when it doesn't, because the thing choosing your model now also sells you the seat.

If your team lives in Copilot, an admin has to switch the MAI-Code policy on before anyone can use it, so decide deliberately rather than drift onto a default. Benchmark it against your current model on one real, messy task before you standardise on it. The broader watch item is how many tool makers follow Microsoft down the stack and start shipping the model, not just the interface.

Four tools to install or test this week. Productivity is carried by the four stories above, so these run marketing, design, sales, and dev.

Marketing: Klaviyo Composer builds the campaign from a sentence. Klaviyo opened Composer in public beta on June 30, an agent that takes a plain-language brief ("winback for lapsed VIPs, email and SMS") and assembles the audience, per-channel copy, and send, grounded in your store data. Customers get free credits to start. Build one real reactivation campaign against your own data and approve it before it goes out.

Design: Figma turns a design layer into live code. Figma's new Code Layers, shown at Config and rolling out through July, convert any layer into an interactive, code-backed element from a click or a prompt, so you can branch directions side by side in the same file. It moves the prototype a step closer to the real thing. Join the waitlist and turn one static prototype into a working code layer.

Sales: HubSpot opens its Prospecting Agent to every paid tier. HubSpot's Breeze Prospecting Agent is no longer early-access only; any paid customer can now let it find net-new contacts and draft outbound against a defined ICP, alongside a shared Breeze workspace for prospecting and meeting prep. Switch it on, have it build one outbound list, and check its picks before you send.

Dev: GitHub Copilot Vision reads your screenshots. Copilot vision went generally available on July 1 across every plan, so you can drop an image or PDF into Copilot Chat and have it reason about a mockup or a broken UI next to your code, no admin toggle. Paste a screenshot of a design or a bug and ask it to generate the matching component.

Where the region and the wider market gather this autumn. Book the ones worth the flight now, 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