Agent 2.0 (beta): employees, photos in chat, sound — and a big reliability wave

We bumped the version honestly: employees, a multimodal chat, sound, and the new reliability make this a different Agent from the one that shipped a week ago. But it's too early to drop the beta badge — a conversational assistant only gets polished on real conversations. Every week we review where Agent misread a brief, where a prompt came out weaker than it could, where an answer ran too long — and we fix it.

So the most useful thing you can do is simply use it. And if Agent answers oddly or misses the point — tell it right in the chat, or write to support. Every such case turns into a concrete fix.

When one head answers for everything, every conversation starts with setup: "I need a marketplace card, vertical, with infographics, bigger text…". An employee has already been through that setup. The Card Designer thinks in slides and offers from the first message, the Videomaker in shot lists and first-second hooks, the Copywriter in keywords and buyer benefits. You just state the task.

The roster follows real demand: we looked at what marketplace sellers and small businesses order from freelancers most, and built employees for exactly those jobs. Each has its own professional techniques and knowledge base — including marketplace requirements checked against the official Wildberries and Ozon help pages.

An employee is not a separate plan or a subscription to a person. It's the same Agent working in its own chat within its specialty: the same tokens, the same limits, the same price card before every generation.

The project chat list gained a "Hire an employee" button. It opens a gallery of role cards describing what each one does. Pick one — and the employee gets its own chat with its own starter suggestions; in the chat list it carries a role badge so it never blends in with regular conversations.

The fast path is a slash command right from the main chat: the specialist gets hired and your brief lands in its new chat immediately — together with any attached files. The dialogue model for each role comes with a sensible default, and you can change it in the same picker as Agent's. Hiring any number of employees costs nothing by itself — you pay, as usual, only for turns and approved generations.

In the previous announcement "photos in chat" honestly sat under "What's next" — that next has now arrived. Attach an image right in the conversation: as a reference ("make it in this style"), as a source ("here's my product — shoot it on white"), or for editing. For precise fixes there's the brush: open the image, select an area, and ask to redo only that part. Materials already uploaded to the project are available through the paperclip too.

Sound arrived through the same flow as images and video: ask for a jingle, a background track for your clip, or a voiceover — Agent picks the model, shows the price card, and launches after your Approve. Voiceover reads your text verbatim; for music you set the mood, tempo, and duration to match your clip.

Replies no longer land as one block after a pause — they stream word by word, like a live conversation. While Agent works you can see what it's doing right now: "Searching materials", "Picking a model", "Composing the prompt" — and who is working, when a specialist runs the turn.

When Agent offers options — "vertical or square?", "which of the three concepts?" — buttons appear under the message. One tap instead of retyping. And new chats now take their name from your first message, so the chat list stays readable.

We went through every place Agent could go silent or stumble — and closed them. No message is ever left unanswered: whatever happens inside, the chat gets a plain explanation instead of an endless spinner. If your dialogue model is overloaded or failing, the turn automatically completes on a fallback — with a visible badge showing which model answered.

Agent also sees its own work better now: ask "what's up with my video?" and it checks the status of launched generations and answers to the point. If a generation fails, it figures out the cause, rewrites the prompt, and proposes a new run as a card — not a blind retry on your dime.

The chat now shows a counter: how many messages you've used today, your plan's limit, and — on the free plan — how much of the monthly trial remains, with a short explainer of what affects the turn price. No trips to settings to know where you stand.

The proposal card got more patient: its lifetime is longer, and if it expired while you were discussing details, one click brings it back — no new paid turn. Type "Approve" as text and Agent highlights the card it refers to, but the launch still happens by button only.

And two honesty rules. First: if you keep approving cheap generations without edits, Agent offers — once — to auto-run the small stuff; your call, and it can be turned off. Second, the firm one: a platform-side failure is not your problem. Such turns aren't charged, and failed paid generations refund their tokens — with the status right in the error card.

Agent can now propose Seedream 5.0 Pro for photorealistic commercial hero images, product photography, or edits to materials, colors, and individual objects guided by annotated references. It is especially strong on natural lighting, skin, and material texture, with 1K/2K output and up to ten references.

This is not a new undisputed image-model leader. Early hands-on tests praise product photography, commercial posters, and local edits, while complex text, infographics, and UI still show errors. Nano Banana Pro remains the choice for 4K brand scenes, and ChatGPT Images 2.0 is the most proven general option for complex instructions and layouts. Seedream 5.0 Pro stands out for photorealistic commercial imagery and cheaper iterations.

The Agent model picker has two new options as well. Grok 4.5 is a fit for bold ideas, lively writing, and an unconventional point of view. Fable 5 is the highest tier for the hardest briefs and long agentic work. As before, every turn price is visible before you choose.

The roster and the abilities keep growing. On the roadmap: