The Year of the Agent is an unusually resilient franchise. It can survive a delayed rollout, an ambiguous definition, and the discovery that the agent needs someone to approve its expenses. The next season is always going to be the one where it really gets going.
Sequoia’s May recap of its April AI Ascent event described the combination of models, tools, and harnesses as central to the agent story. It also acknowledged a gap between emerging capabilities and deployment inside large organizations. The account is an investor’s event recap, not a representative measure of how much work has been automated.1
The organization is not an API endpoint
A company may need to determine who owns the result, which records can be accessed, how errors are corrected, and what happens when the relevant employee leaves. These are organizational questions with technical consequences. They cannot always be solved by improving the model.
Consider a hypothetical purchasing assistant. It can locate a supplier, compare options, and draft an order. The organization still needs a rule for when a price change requires approval, which supplier terms are acceptable, and who is responsible if the purchase was inappropriate.
The model might perform its assigned task perfectly while the assignment itself remains underspecified. This is a surprisingly durable feature of work.
The invisible middle
Product demonstrations tend to show the request and the result. Between them is the material that makes a system useful over time: permissions, retrieval, error handling, evaluation, integration, and clear stopping conditions.
The commercial opportunity may reside in that middle. A company that understands a workflow deeply enough to make delegation reliable can deliver something more valuable than a conversational interface. But it has to do the work, not just rename its interface an agent and rely on the noun to imply the missing architecture.
A cautious buyer can start with a bounded task and compare the result with the existing workflow. That is less photogenic than replacing a department, but it gives both the buyer and the vendor something to check.
A year is a marketing unit
The problem is that adoption rarely respects the appointment. Different tasks become practical at different times. Different organizations face different constraints. A small team trying a low-risk workflow and a large institution handling sensitive records should not be expected to move in lockstep.
Ask what the agent may now do without permission that it could not do before. Then ask who notices when it gets that new freedom wrong.
Keep a record of the interventions as well as the completions. The person rescuing the workflow is part of its operating cost.
The pilot has not yet met the shift change
A pilot benefits from attention. Its users may be unusually motivated. Its builders are nearby. Exceptions can be resolved by people who understand the intent.
A routine deployment has to survive less favorable conditions. New employees arrive. Inputs change. The enthusiastic champion moves to another team. A task occurs at an inconvenient time, and the person who knew the workaround is unavailable.
The companies that bridge this distance may be less theatrical than the ones that generate the best demonstration. They will have opinions about logs and rollback. Their sales materials may contain a disturbing number of details about the customer’s actual workflow.
Renewed on the strength of the work
The next season needs evidence from the workday. Preferably from somebody who did not help produce the trailer.