2025 was the era of agents. Agent builder tools were everywhere (and still are), walk into any conference and you would hear non-stop talks on how to build your own agents.
But looking at what companies actually shipped with them, it boiled down to three main categories:
- 1. Customer-facing support agents - the evolved version of old-school chatbots. Great at handling billing queries, replacing WalkMe/Whatfix flows, and helping users self-serve.
- 2. Internal agents for behind the scenes workflows - These aren’t customer-facing at all. Background agents that sync data across systems, detect anomalies in real time, enrich records, or orchestrate complex batch jobs so the product just works better without anyone ever chatting with them.
- 3. Core workflow automation agents - the ones pitched as a natural language interface to “do anything” in the product. For example, instead of your customer dragging arrows and blocks in your workflow automation product, they would just describe the workflow in plain English to the agent. In fact, many companies went as far as making this their core UX.
But the biggest shift is happening with #3. In recent conversations, I'm starting to notice companies pulling back hard on these core workflow agents.
Most teams are noticing their customers defaulting to Claude Code, Codex, or their preferred coding agent (or cowork product) to interact with them instead of adopting yet another agent. The biggest shifts that have driven this are:
- 1. The underlying models and harnesses have improved dramatically, closing the gap on product context limitations.
- 2. Growing availability of connectors like MCP (and the ecosystem keeps expanding).
- 3. Ease of use, since these tools already carry rich context about the user’s needs and workflows.
Companies are dialing back investment in these standalone “product agents.” The winners will be the platforms that make it dead simple for powerful external models and agents to act meaningfully on your product.
One open question this raises: how do we maintain observability and rich data on how customers are actually using your product, and where they’re getting stuck, when interactions move outside your owned agent and product?