StackOne

There's a limit most developers don't know exists until they hit it. OpenAI caps tool definitions at 128 per request. If you're loading more than that, the request fails. A developer spends an afternoon building out their connector setup, calls the API, and gets an error that doesn't obviously explain what went wrong. That's usually the moment teams discover that tool loading doesn't scale. The workaround is manual curation — maintaining a list of which tools to load for which workflows. It works until you add a new connector. Then the list needs updating. Then again for the next connector. Six months in, it's load-bearing infrastructure that nobody wants to touch. Tool Discovery removes the list entirely. The agent searches for what it needs at query time — no static config, no 128-function ceiling, no maintenance burden that grows with your stack. Docs here: https://lnkd.in/e6wTi_T8.

We talk to a lot of teams building production agents. Once they've connected more than a handful of platforms, they almost always hit the same moment: every platform they connected brought more tools and, if you add enough of them, the context window fills before the agent even reads the prompt. The agent didn't error. It just stopped finding the right tool, started skipping steps, or called an action that was close, but not the one it needed. The cause: loading every tool definition on every request. At 200+ tools, that's up to 200,000 tokens of context devoted to tool descriptions before the agent reads a single word of your prompt. Tool Discovery is built for this. Instead of loading the full catalog, the agent searches for the right tool at query time and executes it directly. Two meta-tools replace hundreds of definitions. Context drops from 200k to 8k tokens. First-try accuracy: 92.8% — including connectors the model was never trained on. At 44,453-tool scale, StackOne scores 54.4% vs. Anthropic's 30.0%. One URL parameter to enable it. Swipe to see how it works. https://lnkd.in/et8UYC4A

Somewhere between 50 and 200 connected tools, something breaks. The agent doesn’t crash, it just starts making wrong choices. Calling the wrong tool and missing the right one, quietly degrading on tasks it handled fine last week. You spend days trying to figure out why. Eventually you realize it’s not the model. It’s the context. Loading 200+ tool definitions on every request burns through an LLM’s attention before it even reaches the task. OpenAI caps at 128 functions. Most teams hit this wall and resort to maintaining a static list of tools per workflow — a workaround that creates its own maintenance burden with every new connector added. Tool Discovery is our answer. Two meta-tools replace the full catalog: tool_search and tool_execute. The agent searches semantically for what it needs, runs it server-side through StackOne, and overhead stays constant no matter how many systems you connect. 94.9% nDCG@5 on single-tool retrieval. 54.4% nDCG@10 across 44,453 tools, vs. Anthropic Tool Use at 30.0%. Built on a fine-tuned bi-encoder, scoped to your connected stack. Available today. Learn more: https://lnkd.in/dWnptqxH

We're officially in town for SaaStr Annual and LangChain Interrupt this week. If you're building AI agents and hitting the wall on reliable tool connections, come find us. We'll show you what production-ready actually looks like. SaaStr Annual — May 12-14, San Mateo LangChain Interrupt — May 13-14, SF DM us to lock in a time, or just grab us between sessions.

StackOne offsite 2026 ✅🌍 just back from an incredible week in Croatia 🇭🇷 With a team spread across the globe, our time together in person is invaluable. A week of collaboration, hackathons, exploring, karaoke, and lots of good food. More than anything, it was a reminder of the brilliant people that make up this team! ✨

We're hosting a morning run during SaaStr + Interrupt week. May 13 · 7am · Embarcadero, SF — two paces, coffee after. No conference badge needed. Come run with us → https://lnkd.in/d4pxyi_Y

The StackOne team is heading to SaaStr Ai Annual next week 🚀 (12-14 May, San Mateo). If you're building enterprise grade AI agents, the integration layer is where things get hard. 🛠️ That's what we work on. Come find us between sessions and we'll show you what production-ready looks like in practice ⚡ DM Romain Sestier, Alexander Cox, Elena Angela or Shawnee Chase Foster to grab a slot, looking forward to meeting you there 👋 #SaaStrAnnual #EnterpriseAI #AIagents #StackOne

The playbook for context engineering in AI agents is being written right now — by the teams actually building. Context windows, tool use, memory, retrieval, MCP, agent harnesses. The discipline of designing what information an agent has access to and how it uses it is the core problem no one has fully solved yet. The real question isn't which model or framework to use — it's what your agent can see, what it can reach, and how you keep that scoped and efficient at production scale. Romain Sestier will be in San Francisco on May 5th for Context is King #4, demoing the StackOne integration layer alongside engineers from companies building at the frontier. Hosted by Flow AIand Aiven. If you're in the Bay Area building production agents, this is worth the evening. Sign up here: https://lnkd.in/emWxVCuN

Indirect prompt injection is now OWASP's #1 threat to LLM applications, found in 73% of production deployments. The attack is simple: hide instructions inside an email, a CRM note, or a support ticket. The agent reads the data, sees the hidden command, and follows it. No clicks. No user interaction. Our CTO Guillaume Lebedel dug into the research — and the numbers aren't reassuring. The ICLR 2025 Agent Security Bench found 84% of tested agents vulnerable. Mixed-type attacks hit 100% success. Even Claude 3.5 Sonnet, which blocks 93% of known patterns, falls to 81% of novel ones. His key insight: defense can't live in model training alone. Models learn to resist known patterns, but novel attacks consistently break through. Defense has to happen at the tool boundary — scanning every tool response before it enters the context window. Guillaume walks through how we built a two-tier defense (regex + MLP classifier) that scans tool results in ~11ms on CPU, hitting 90.8% F1 while staying 81x smaller than the nearest competitor. Read the full piece on our blog. 👇 https://lnkd.in/e2DCu9mm #PromptInjection #AIAgents