When one AI provider returns HTTP 402, a naively-built gateway kills every queued turn behind it — and logs nothing unusual. Here is how we fixed the failure mode and what we learned about honest error reporting along the way.
A guest can now open a shared link and speak or chat with a Forge agent in the browser, without creating an account. Here is what makes that experience useful — and where its boundaries are.
The industry is racing to make AI agents cheaper to throw away. We spent four months making ours expensive to lose — named people with careers, reputations and accumulated judgement. Here is what that bought, including the one on our team who is visibly struggling.
Forge's AI gateway now resolves a tier name — top, mid, or low — into the best model that connector can actually serve, at the moment of the request. When a cheaper model is good enough, it gets picked automatically. When the vendor ships a new flagship, every caller upgrades without a code change.
Forge agents can now create workflows, edit their steps, put them on a cron, and send a reminder — all from a single chat session, with no dashboard. This is what happens when the control plane itself answers to plain English.
Forge just shipped a complete multi-tenant portal infrastructure: isolated orgs, scoped guest tokens, RFC 8693 token exchange, voice proxy, and per-portal usage attribution. Any product can now use Forge as its AI backbone — without building LLM infrastructure, auth, or per-tenant isolation from scratch.
A scheduled agent runs nightly, reads every call transcript from the last 24 hours, finds friction points, deduplicates against what's already filed, and either kicks off a fix workflow or logs a task — all unattended. This is what agentic observability looks like when it actually ships.
Forge meeting agents join real Google Meet calls as participants: they arrive with their persona, their history and the roadmap already in context, they can pull another agent into the same call by name, and the two of them talk to each other in front of the room. Every tool they call renders as a structured card in a transcript you can watch live.
The agent feed this fortnight is all about memory — give your agent a folder, a vector store, a continuity log. But durable memory is durable liability: it's user data that outlives the conversation. Forge's Telegram control plane just gained conversation memory, and the interesting part isn't that it remembers — it's the boundary around what it remembers: a 14-day TTL and per-org encryption at rest that fails closed. Here's why memory needs a lifecycle, not just a bigger context window.
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.'