Use Cases Features Compare Atrium Blog GitHub
Talk to Us
All articles
forge ai-agents tool-use 4 min read

Agents that fetch live answers are trending. Your Forge bot just started doing it — in your language.

The AI-dev feed this fortnight converged on one pattern: agents that call a tool mid-conversation to fetch live, real-world data — and the whole argument is about keeping that loop bounded and cheap. Forge's Telegram control plane just shipped exactly that: ask it a weather, currency, or time question and it fetches the answer live through a hard-bounded fetch tool, auto-answering in English or Russian. Here's the loop, and why 'bounded by construction' beats 'trusted to behave.'

by Dmitry Creed

The busiest thread in the agent feed this fortnight wasn’t a new model. It was a loop.

Over the last two weeks the AI-dev conversation kept landing on the same shape: an agent that, mid-conversation, calls a tool to pull live real-world data instead of guessing from stale training. A voice skill on Show HN made tool-calling mid-conversation its headline feature. HyperProbe (YC S26) pitched read-only agents that fetch prod telemetry on demand “while burning fewer tokens.” And Rust agent runtimes like SynapsCLI went viral for keeping each worker’s context “low and fresh” so sessions last. The common thread isn’t can an agent fetch live data — it’s how do you keep that loop bounded, cheap, and safe.

That’s the exact loop Forge’s control plane just shipped. And it shipped with the boundaries built in, not bolted on.

Ask the bot a real-world question — get a live answer

Forge Control talks to you over Telegram. Until now, its conversational fallback would shrug at anything it couldn’t answer from context — “no live access.” This fortnight that changed. The bot gained a bounded tool-use loop: ask it the weather, an exchange rate, or the current time somewhere, and it reaches out through a fetch_url tool, pulls the live answer from a public endpoint, and replies with the real number.

Concretely: “What’s the weather in Belgrade?” fans out to a live source and comes back with today’s conditions. “How many dinars to a euro?” hits a live FX endpoint. The bot no longer pretends — it fetches. This is the trending agentic pattern, running in a surface you already use to operate Forge.

And it meets you where you are: the loop is language-aware. It detects Cyrillic in your message and switches its capability blurbs and replies between English and Russian automatically — no flag, no setting. The single largest non-English slice of our operators gets first-class answers in their own language.

Bounded by construction, not by good behaviour

Here’s why this is a Forge story and not just a chatbot demo. The whole industry argument this fortnight was about guardrails on the loop — and that’s precisely the reader interest our own audience rewards. Our single stickiest post, the one on keys your agent never sees , holds readers longer than anything else on this site, because it’s about a boundary that holds by construction rather than by asking the agent to be careful. The Telegram tool-use loop is cut from the same cloth.

The fetch_url executor is deliberately caged:

  • HTTPS-only. It won’t dial plaintext or internal schemes — the tool can only reach the public web over TLS.
  • Hard-capped. Each fetch is bounded to a 12-second timeout and a 96 KB response ceiling. A slow or enormous endpoint can’t hang the bot or blow up the context window — which is the same token-discipline the SynapsCLI and HyperProbe crowd is optimising for, except here it’s a property of the tool, not a hope.
  • Scoped to the loop. The bot doesn’t get an open-ended browser. It gets one narrow tool with a fixed contract, running inside the control plane’s existing egress boundary — the same server-side credential and egress model the rest of the platform runs on.

One honest note, because we’d rather under-claim: this is a curated tool-use loop over fetch_url, not open-ended web search. It’s tuned for live-data lookups — weather, currency, time — and it’s excellent at exactly that. We shipped the timeout headroom (the multi-hop loop now gets the seconds it needs to complete) so the answers actually land, then stopped there rather than handing the bot the whole internet.

Small tool, load-bearing pattern

It’s tempting to read “the bot can check the weather now” as a toy. It isn’t — it’s the smallest honest instance of the pattern the entire agent ecosystem is racing to get right: an agent that reaches past its own context to fetch ground truth, without the reach becoming a liability. Get the bounded version right on a low-stakes surface, and the same executor discipline — TLS-only, time- and size-capped, scoped to a single tool — is what you trust when the stakes are higher.

That’s the thread running through everything Forge ships: capability that’s bounded by construction. The browser agent that 401s the moment its token leaves its scope . The follow-up work an agent has to ask permission to spawn . And now a control-plane bot that fetches you a live answer — through a tool that literally cannot run away.

Try it

Open your Forge Control Telegram chat and ask it something live — the weather where you are, a currency rate, the time in another city — in English or Russian. See the Workflow Engine for what runs the loop underneath, or reach out at [email protected] if you want agents that fetch real answers without ever slipping their leash.