Future AGI

55 tools are useful. Introducing all 55 before the first task is a billing event. In our last webinar with Snowflake, Josh showed a cleaner coding-agent harness: keep the full tool catalog available, but start with two tools - one to find what the agent needs, one to invoke it. Yes, that adds a lookup. With a large registry, it’s a trade-off worth measuring against loading every definition up front. Watch the clip. Full webinar here- https://lnkd.in/gU9p-jrN

“Give the agent more context.” “Keep its context lean.” Both pieces of advice are right, which is annoying. Rishav asked Josh where the line actually is and the distinction is worth the minute. Full Future AGI × Snowflake webinar: https://lnkd.in/g-qZpFNh

A hijacked agent doesn't throw an error. It finishes the run, just on a different task: the one an attacker planted in a PR title or web page it read along the way. So a test suite that asks "did the agent finish?" passes. The only evidence is the tool calls, if you logged them. We wrote up how to test for this: four disclosed incidents from 2025–26 and a three-layer setup of prompt-injection red-teaming, tool-call evals, and runtime guardrails 👇  #AIAgents #AISecurity

A claims voice agent that resolves 41% of calls isn't saving an insurer much. For most callers, it's just an extra step before a human. That's where a leading insurer we worked with was, four months after launch. Most of their ~40,000 monthly claims calls ended in a transfer or a hang-up, so they paid for the AI and then for the adjuster who took over. It also puts renewals at risk: a research firm found 52% of customers with a poor or "just OK" digital claims experience are likely to leave or not renew. Their dashboards showed calls failing, but not why. Future AGI helps teams find exactly where agent conversations break, and prove the fix before customers hear it. Here's how: --> 𝐒𝐢𝐦𝐮𝐥𝐚𝐭𝐞. 2,400 calls modeled on their real claims line, with calm, stressed, confused and impatient callers. 44% were resolved, close to production's 41%, so the simulation could be trusted. --> 𝐄𝐯𝐚𝐥𝐮𝐚𝐭𝐞. Four evals scored every call. Did the agent remember what the caller said? Explain next steps without promising coverage? Avoid loops? Hand off when asked? 984 calls failed at least one. --> 𝐆𝐫𝐨𝐮𝐩. Error Feed turned those failures into four root causes. The surprise: the agent's "safe" non-answer to coverage questions lost more than three times as many callers as risky promises did. --> 𝐅𝐢𝐱 𝐚𝐧𝐝 𝐩𝐫𝐨𝐯𝐞. Four targeted changes with no model change, each re-tested on the same 2,400 calls. Six weeks later it was fully rolled out, and sixty days after that: → 68% of calls resolved by the agent, up from 41% → 6,400 fewer transfers a month → Hang-ups down from 21% to 10% The same approach works for agents that quote, qualify leads or handle renewals. If you run one, we'd be glad to test it against your real call mix. Swipe through for the full breakdown, including one call scored turn by turn (our approach starts from slide 6) 👉

Things nobody tells you about coding agents on data tasks: — SQL that runs isn't SQL that's right — your agent loads 30 tools to use 4 — one SELECT * can dump 13,000 rows into context — you're paying for "USD" 13,000 times when once would do — your AGENTS.md gets ignored once the task runs long. that's a context problem, not a prompt problem — more business context helps. more tools hurt Josh (Snowflake) and Rishav walked through all of this, with a live demo on data-eng-bench. Recording's live 🎥 : https://lnkd.in/gwT8kPB2 Repost if your agent has ever been confidently wrong ♻️

Future AGI is now available on Google Cloud Marketplace. AI teams across industries already rely on Future AGI as the self-improving layer for their agents in production. With this listing, organizations building on Google Cloud can add it to their stack the same way they adopt the rest of their infrastructure: through their existing Google Cloud account and through a procurement process their teams already know. For us, this is about reach. Many of the teams building the most ambitious AI systems today do it on Google Cloud, and we want Future AGI to be available wherever that work happens. Grateful to the Google Cloud partner team for their support in getting here. #GoogleCloud #AIAgents

A voice agent can get a perfect transcript score and still deliver a terrible call. That is the problem with treating voice evaluation as chatbot evaluation with an audio file attached. The words may be right while the caller was interrupted twice, an account number was heard incorrectly, or the agent took just long enough to make them say “hello?” Voice needs its own evaluation stack: audio quality, speech recognition, model behavior, voice output, and the conversation as a whole. When a call fails, you need to know which layer owns it. We wrote a practical guide to what to test, instrument, and trace before they become someone’s least favorite phone call.

If coding agents and data workflows are already on your roadmap, this is a good chance to hear how another experienced builder thinks about the problem and compare notes with people working through it too. AI engineers, data teams, platform engineers, and technical leads are all welcome. Register here: https://luma.com/uktqv4zk