The day AI agents met the plumbing around them

Every announcement worth reading today circles the same shift. AI agents are no longer a demo on a slide; they are running in production, touching data that people and customers depend on, and the hard problem has moved from whether an agent can do the work to what you put around it so that a mistake, a runaway cost, or a missing piece of context never reaches the customer.

That is a useful lens for anyone choosing tools this quarter. The interesting part of a product is increasingly not the agent itself but the plumbing: the copy that saves you when something goes wrong, the dashboard that tells you an agent is failing, and the connections that let one agent see what another already knows.

Plan for the agent that deletes something you need

A human who deletes a record usually remembers doing it. An agent working across many records at speed can remove something important before anyone has looked at a screen. The lesson is not to trust the agent more. It is to assume that at some point an agent will destroy data you cannot rebuild from memory, and to have a copy you did not have to plan for in the moment.

SaaStr described exactly this from its own operations, writing that "We run SaaStr with three humans and 20+ AI agents in production" and recounting two separate occasions when an agent wiped something the business could not run without [1]. Both times a copy that nobody had deliberately created is what brought it back. For a business evaluating any tool that lets agents write and delete, the question to ask is simple: when an agent removes the wrong thing, how do we get it back, and how long does that take. If the honest answer is "we hope it does not happen", you do not yet have a plan. This is the practical side of knowing what happens to your data when you leave a tool — the same export and recovery habits protect you from the agent inside it.

Watch the agents you put live

One agent you can keep an eye on. Ten agents, built on different platforms by different teams, you cannot. At that point you need the same thing an operations team needs for any production system: continuous monitoring that tells you when something is failing or costing more than it should, before a customer feels it.

Salesforce moved in this direction today. Its Agent Fabric update is built around the idea that "New Agent Evals gives leaders the confidence to scale more agents into production by continuously evaluating agent performance and costs, no matter what platform the agent is built on" [2]. Two words there matter for buyers. "Continuously" means you find out about a problem while it is happening, not in a monthly review. "No matter what platform" means the monitoring sits above any single vendor's agents, which is the shape you want if you are going to run agents from more than one supplier. The deeper point is that an agent's cost and failure rate are now operational metrics, not afterthoughts. Deciding which actions an agent is allowed to take in the first place is a related discipline, and one we have written about in what automation should never decide.

An agent is only as good as what it can see

The limitation most teams hit with AI is not the model. It is that an insight produced in one tool stays there. The next person, or the next agent, cannot act on something they cannot reach. Salesforce put a number on this today, reporting from a survey that "As more and more businesses rely on AI, workers are experiencing a bottleneck: interface isolation" [3]. When insights go unshared, the organisation pays for the same thinking twice and acts on less than it knows.

This is worth understanding before you buy. A tool that produces brilliant AI output inside its own window, with no way for the rest of your stack to use that output, may leave you no better off than before. The value is in the sharing, not the generating.

Carrying context across the tools you already use

If isolation is the problem, bridges between tools are the response, and two vendors shipped bridges today. Dropbox announced a way to "Bring trusted Dropbox project context into Microsoft Copilot to find information, create across Microsoft 365 applications, and keep work connected and governed in Dropbox" [4]. The phrase to hold onto is "connected and governed": the content becomes visible where you work without leaving the place where its permissions live. Zoom made a similar move, describing how "ZoomMate and Zoom's Unified Context Layer help close the gap between enterprise conversations and the outcomes that matter" [5] — in plain terms, making what was said in a meeting usable by the systems that act on it afterwards.

The trade-off to weigh is governance. The moment one tool can read another's content, you have to be sure the permissions travel with it, so that a bridge does not quietly widen who can see what. That is the right question to put to any integration, and it is the quiet reason why your tools do not talk to each other in the first place: connecting them safely is harder than connecting them at all.

The data underneath the agent

The last thread is the least visible and arguably the most important. An agent is only as reliable as the data it reads, and data loses accuracy at every step between where it is created and where it is used. SaaStr covered Sumble, a data company whose work speaks to exactly this gap, noting that "Amelia called Sumble the best company you probably haven't heard of yet" and that its customers include Snowflake, Databricks, Figma and Wiz [6]. The headline figure in the piece — that data which is accurate at the source can be far less accurate by the time an agent acts on it — is the point every buyer should sit with.

It means a vendor's accuracy claim about its data is a claim about the source, not about what reaches your agent after enrichment, matching and transformation. If the pipeline loses accuracy at each hop, the number at the end is the one that governs your results. We have written separately about why data gets out of date, and the same discipline applies here: treat accuracy as something you measure where the decision is made, not where the data was born.

What ties the day together

None of today's announcements is about making an agent smarter. Every one is about the systems around it — the backup, the monitor, the shared context, the data quality. That is a healthy sign for buyers. It means the market is starting to treat AI agents the way it treats any other part of production software: something that will fail, cost money, and depend on the quality of what you feed it. Choose tools that take those four things seriously, and you will be choosing well.