Adapted from an internal post on 1 September 2026.
Three weeks ago we made two commitments: customers first, and craft over speed. Since then we have spent hours on calls with the people building agents in production: implementation partners, vertical AI companies crossing from internal use to general availability, and operators who need one agent to work. This letter is what we heard, what we shipped because of it, and where it is going. Our goal is for Sapiom to be the home for the agents that run your business.
What we heard
The hundred teams we spoke to span fintech, luxury retail, insolvency, revenue operations, marketing, healthcare, frontline HR, government contracting and corporate travel. There is no overlap in vertical. The pain is nearly identical.
People cannot tell whether a run worked. A run reports success and returns nothing usable. Six runs, forty-five failed fetches underneath, all six green.
There is no single surface for cost, failure and activity. Spend lives in one dashboard, alerts in Slack, error rates somewhere else. One team told us they have five hundred gigabytes of execution logs and “there needs to be an observability which we don’t have right now.”
Everyone wants to own their orchestration, but they end up rebuilding the same platform layer. A founder who has tried about a hundred architectures wants a partner “one level up.” A frontline-HR platform onboarding 150,000 workers a month: “everyone has to build the same infra. I’ve talked to so many of them.” They build custom LangChain abstractions, but are struggling to hire the platform engineers to extend them.
Scale and cost economics are the looming problem. Most raised it unprompted. Their agents will be measured, as one CEO put it, “ultimately by the business outcome they deliver,” and the cost of that outcome has to be visible per run.
What we shipped
You build a fleet, not an agent. Every team we spoke to is running several agents that call, signal and hand off to each other. We rebuilt navigation around projects and introduced the agent map so you can see the agents in a system and the links between them, then trace through a run.
App Links: put an agent behind a URL. Every agent you build can publish a shareable app dashboard, published straight from the MCP. Share it across your organization, make it public with a daily spend cap and wake rate limits, or give it a static endpoint so a webhook can wake it. You pay only for what runs.
Connectors. Notion joins Linear as a one-click OAuth connector, so agents attach to the workspace you already have. Google Drive and Gmail are next.
Triggers and events. Public webhook ingress with signed, rotating secrets. A tenant events API. Manual replay for failed fires. Agents now start from the outside world, not only from a schedule.
We’re using agents for GTM and SalesOps. Every customer and discovery call is now processed and cleaned by one agent, sent to another for analysis, and to a third that updates our Notion and CRM. They execute automatically from external webhooks and our event patterns. It took an afternoon.
We open-sourced the Agent Studio because building agent fleets is universal. Regardless of your framework or your hosting, you still need to design the system, debug errors, version and test, and observe agent interactions. We aim to support any framework and integrate with every platform.
Where this goes
These features lay the UX foundation for the work we are doing with the upcoming Sapiom Agent Builder. What we’ve all found is that foundational models are bad at building agents, especially on the first go. It’s exactly like building any non-trivial project with AI: if you expect magic, what you get is slop. We’re creating the Sapiom Agent Builder to solve that. This is the IP entrypoint that shows how highly capable agents can be built reliably and securely on Sapiom, without sacrificing intelligence. We are launching Sapiom Labs to publish benchmarks, research, and blog posts that prove the quality of the agents we craft.
There is a lot more that deserves honorable mention; here are just a few. The SDK shows step-level input and output, for clearer debugging. Connectors are in flight, so agents can connect to the user’s existing workspace. The webapp billing page shows users their capability usage and plan limits.
Sapiom agents have boundless capability, and in the weeks post-launch we have focused on making them more observable, reliable, and intuitive for new users. A warm welcome to Brett and Shang; it’s a great time to join.
Looking forward to what’s next!
— Dave
P.S. Curious what your agent fleet would look like on Sapiom? We’ll show you.