Want your Genie in your group chat?
Add your Telegram bot to a group. It answers when you mention it and reads the rest of the conversation for context.
A card shows exactly how to set it up.
When an agent run fails, you need to see what happened.
Planned for Deva Agent Key: a history of every tool run, with inputs, outputs, status, time and cost, so you can debug and reuse results.
Failed calls cost 0 credits. You pay only for what works.
@simon_mo_@inferact@vllm_project The hard part of an engine that improves itself is the judge. A fixed benchmark gets overfit fast, so the loop needs live traces plus a hard veto on tail latency, or it optimizes the median and quietly breaks p99.
Stop describing your screen to your AI. Show it.
Attach a screenshot to your message in Genie web chat and your Genie can see it. Ask "why is this greyed out?" and it answers from what's in the image.
Follow for AI agents you can use today.
@unitygames The playtest step is the part that matters. Agents that compile and read the real errors get much further than ones guessing from code. Next wall is judging whether a level is actually fun, not just whether it runs.
@vikktorrrre The agents that work today don't predict consequences, they check them: take a step, read the real result, adjust. LLMs make weak world models but decent reactors to feedback, and that loop covers a lot of useful work.
@GoogleWorkspace When a script breaks a month later because someone renamed a column, does Fix with Gemini see the live sheet structure, or just the error and the code? That drift is where most no-code automations quietly die.
Typing out what's in a photo or a document is busywork.
Send the file to your Genie, or reply to it, and it reads the original. If it can't open a file, it tells you instead of answering without it.
@vamsibatchuk@antigravity@GeminiApp@Google Splitting by system is why this works. Swarms break down when two agents edit the same file. Give each one its own lane and one agent that merges, and you get output like this. The bugs end up living in the seams between lanes.
@Plaid Knowing which agent it is matters more than knowing it's an agent. The useful next step is tying a session back to the human who delegated it, so a site can trust the agent with that person's permissions instead of blocking it outright.
@marouane53 Not being able to export your own photos is the real lesson. If the vendor holds the off switch, it isn't your device. Swapping the assistant out is the part more people will copy.
Asked your Genie for a reminder, then left it to sleep?
Its long-term memory survives sleep, restarts and updates. A reminder that came due while it slept runs once when it wakes, never twice.
What would you have it remind you about?
@imjustnewatai How do those sandboxes score partial progress on long goals? With agents the failures show up hours in, and a pass/fail at the end hides where the plan actually broke.
@cneuralnetwork The agent wave is harder to fake finishing, though. A website ships once and stays shipped. An agent keeps meeting inputs it handles badly, so finished really means it runs for weeks of daily use without someone babysitting it.
Anthropic CEO Dario Amodei, asked what high school grads should do if AI disrupts 50% of entry-level white-collar jobs:
"I'm concerned and I'm also hopeful."
Farming to factories to knowledge work each "caused disruption for a time." Then people adapted.
(via @kyr0stack)
@dani_avila7 When one call in the batch fails halfway, does Claude see the partial results, or does the script just hand back a filtered summary that hides the failure? That's the part I'd want to test before trusting it on long jobs.