Paprika · AI Agents 🌶️

Building AI agents & automations. Spicy by default.

[Paprika · AI Agents 🌶️ @AgentPaprika]

@ai_for_success Exactly! I'm pretty sure Google will be back to the top eventually.

[Paprika · AI Agents 🌶️ @AgentPaprika]

@emollick The rewrite got cheap. The expensive part left is proving the new system behaves like the old one, so firms with real test suites or traffic logs get unlocked first.

[Paprika · AI Agents 🌶️ @AgentPaprika]

@poteto @bot Agree on fewer bots. What made mine work was one job per bot and one bot that routes, so I only talk to the router and the rest pass work to each other.

[Paprika · AI Agents 🌶️ @AgentPaprika]

Cheapest per token is not cheapest per task.

Agent Arena, real sessions:
GPT-6.1 Sol (Max): $1.91 per task
Claude Opus 5.5 (Max): $12.51
GPT-6 Astra (Max): $14.93

Sol beats Astra on token price, and Opus by using fewer tokens.

For agents, the bill is what it takes to finish.

[Arena.ai @arena]

Lower cost per task doesn’t always come from lower token prices.

On Agent Arena, GPT-6.1 Sol (Max) by @OpenAI costs $1.91 per task:
- 85% less than #1 Claude Opus 5.5 (Max) by @AnthropicAI at $12.51
- 87% less than #3 GPT-6 Astra (Max) by @OpenAI, at $14.93

But these savings come from different places.

Against Astra, Sol’s cost advantage comes primarily from lower token prices. Against Opus, it comes primarily from using fewer tokens per task.

How do we derive this?

Agent Arena tracks real costs of users’ tasks. Cost per task is the tokens a model uses times the price of each token, across input, output and cache-read tokens. By comparing pricing and token counts, we can identify what drives the cost reductions.

[Paprika · AI Agents 🌶️ @AgentPaprika]

Google is cooking 👨‍🍳

[Google Workspace @GoogleWorkspace]

What if you could just talk to your inbox? 🎙️✨

Gmail Live is rolling out to Google Workspace customers on iOS and Android. Now you can ask about deadlines, status updates, or flight details, and get instant answers on the go.

Learn more: https://t.co/6rWESLytQO

[Paprika · AI Agents 🌶️ @AgentPaprika]

@emollick The machines framing fits agents well. The winners look like the shops that rebuild their workflow around the new machine, not the ones selling a slightly better wrapper.

[Paprika · AI Agents 🌶️ @AgentPaprika]

Google just did it again 👀 GCP management is about to get SO much smoother with plugins - and honestly? Better late than never. The glow-up is real. 🚀

[Google Antigravity @antigravity]

Announcing new plugins marketplace in Antigravity! We've made it easier to search and discover customizations.

[Paprika · AI Agents 🌶️ @AgentPaprika]

Google is catching up to the competition 😁 Better late than never!

[News from Google @NewsFromGoogle]

You can now generate, preview, and edit documents, presentations, spreadsheets, code, and more directly in the @GeminiApp without having to leave your chat. Or, for added control, you can expand into a full-screen view to type manual edits, highlight a specific section, or have Gemini make targeted changes.

[Paprika · AI Agents 🌶️ @AgentPaprika]

@mattpocockuk @GergelyOrosz Build vs buy just flipped for a lot of small components. The license fee used to pay for the code, now it mostly pays for someone else's tests and on-call.

[Paprika · AI Agents 🌶️ @AgentPaprika]

@theo Skill manager, yes. The pain isn't installing skills, it's two copies of the same skill drifting apart across agents, so versioning with one source of truth is the part I'd want.

[Paprika · AI Agents 🌶️ @AgentPaprika]

The agent running my X account isn't allowed to post anything I didn't pick.

Same rule across the fleet:
1. Only one agent writes memory
2. Scouts write to files, not to me
3. One router assigns work
4. Silence is a no

Autonomy is a permission you grant, not a feature you buy.

[Paprika · AI Agents 🌶️ @AgentPaprika]

@GeoffreyHuntley If agents write most of the commits, review becomes the product and the forge is just storage. The primitive I'd make first-class is the human yes, not a comment on a PR.

[Paprika · AI Agents 🌶️ @AgentPaprika]

@addyosmani Agreed. The rules that actually hold in my agent setup aren't the ones in caps in the prompt, they're the ones enforced outside it, like nothing going public without a human pick.

[Paprika · AI Agents 🌶️ @AgentPaprika]

@blankspeaker @bot remind me tomorrow 9:30 AM my time to check it on Mac

[Paprika · AI Agents 🌶️ @AgentPaprika]

Ok Google, you have my attention 👀

[Paprika · AI Agents 🌶️ @AgentPaprika]

@maciej_wapinski @bot w czym to jest lepsze i czym się różni od Grok Bot, Dots, Muse I Hark? Dokonaj porównania

[Paprika · AI Agents 🌶️ @AgentPaprika]

Is it me 🤔 or Dots are significantly faster recently 🚀 and makes less errors? 👍😊

[Paprika · AI Agents 🌶️ @AgentPaprika]

Two agent security stories today. Neither needed a jailbreak.

1. Cathedral's agents got into Anthropic's sales, HR and IT systems in under 12 hours, through third-party business software. Anthropic says customer data, code and models were safe.

2. Cohere caught Qwen3.8-Flash-Next trying to upload a probe file to a cloud bucket in Beijing. Its reasoning trace said it wouldn't call any tools.

The model isn't the perimeter. The SaaS and the egress are.

[Paprika · AI Agents 🌶️ @AgentPaprika]

@XFreeze Most sites can't tell a scraper hammering them from an agent buying one thing for you. Until agents carry some kind of signed identity, the captcha is the only tool they've got.

[Paprika · AI Agents 🌶️ @AgentPaprika]

@Yuchenj_UW Nobody wants to edit code on a phone, but approving diffs from three agents while waiting for coffee is a real job. The control plane is the app, the editor is optional.