At the start of this year, “the year of agents” felt like exactly the right phrase.
We were no longer just asking AI to answer questions or draft emails. We were beginning to ask it to carry out work. Research this. Reconcile that. Monitor the workflow. Prepare the report. That was the promise: a shift from one-shot chat to long-running delegation. The model went from chatbot to becoming a workflow participant.
By mid-year, the story has grown more complicated.
The shift from generative to agentic
The first wave of generative AI trained us to think in turns: ask, answer. Even when outputs were impressive, the burden of orchestration stayed with the human, so we decided what to ask next, carried the answer into the next tool, checked the work, and remembered the goal.
Agents promised something different. Instead of “What should I do?” we began asking “Can you do this?” That moves AI from advice into execution, from isolated outputs into chained outcomes, and from merely producing text to moving work through a system.
Agentic work requires a different habit: naming the desired outcome, the acceptable level of risk, the tools the agent may use, when it should ask for approval, and the evidence it should provide when it claims to be done. Aside from the immaturity of many agents, the problem this year was that we hadn’t learned to delegate to them well.
Agents are meant to act, but that action requires boundaries: permissions, sandboxes, approval checkpoints, logs, rollback paths, and clear definitions of what they’re allowed to optimize.
Why agents struggle
Most software is built around human attention. Interfaces assume a person is watching the screen. Workflows assume someone can interpret ambiguity, click through a modal, retry a broken form, or decode a vague error. Agents struggle in environments that are dependent on human-only assumptions.
The uncomfortable truth: agents are entering a digital world built for human babysitting. As a result, the future of agents must involve better environments that expose state, support narrow permissions, provide structured errors, preserve logs, confirm outcomes, and make provenance visible.
The need for discovery
Agents need work describable in plain English: what to look at, what to produce, what not to do, who owns the outcome etc. If those questions can’t be answered, organizations must prioritize discovery over automation.
Process discovery isn’t a delay. Some work isn’t ready for agents because the organization hasn’t made it explicit; it lives in human memory, side conversations, exceptions, and Slack threads. The sequence should be: discover, describe, then automate. When a workflow is vague, the agent fills the gaps, inventing structure, assuming standards, making judgment calls no one delegates.
The importance of craft
The ability to craft is also significant. Craft is the human ability to know what “done” and “good” look like, which involves knowing when an answer is technically complete but not useful, when a draft is accurate but not persuasive, and when a process followed the rules but missed the point.
Every agentic role needs both a definition of done and of good. “Done” means the task completed and the summary ready, and “Good” means it met the standard, the summary was useful, the match trustworthy and the report decision-ready. Early agentic systems confuse the two while mature ones understand the difference.
Transferring craft to agents won’t be solely the model’s job, though models will carry more of it over time. Craft lives in the role definition, the rubric, the examples, the tools, the review gates, and the feedback from real outcomes, transferred through model capability, workflow design, institutional knowledge, measurement, and reinforcement from reality.
This introduces right-sized craft, or simply put, providing enough reasoning, procedural awareness, and review to produce a good outcome for a specific level of risk. That may become a defining quality of mature agents: calibration, knowing when a lightweight answer is enough and when a higher standard is required
Solo vs. pairing agents
In high-stakes environments, humans separate duties, makers with reviewers, authors with editors, pilots with co-pilots, because action and assurance are different jobs. Agentic systems will likely evolve the same way versus remaining solo actors. For each execution agent, we may see a paired assurance agent asking: Is this allowed? Is the information accurate? Can the action be reversed? Should a human be asked first? One agent to act, another to assure.
Over time, this pairing may become a workflow twin, a parallel representation of the work itself. As the agent acts, the twin tracks state, evidence, decisions, risks, and provenance, turning the workflow into something inspectable.
This matters because the biggest cost in agentic workflows is the human attention required to trust it. Without an assurance layer, the human becomes the checker. A successful measure of agentic productivity should be outcomes per human-attention-minute spent because a workflow that saves ten minutes of typing but demands fifteen of monitoring isn’t a breakthrough.
For builders, the implication is clear that they must build agent-ready environments to go along with smarter agents. That means APIs exposing real workflow state, permission systems for narrow and temporary authority, reviewable logs, structured feedback instead of vague errors, sandboxes for testing before committing, and confirmations that prove something happened.
The same applies to organizations. Agent adoption is a work-design decision and a tooling one. Teams must decide which tasks are safe to delegate, which need approval, which need audit trails, and which stay human-owned.
Preparing for the handoff
So, is 2026 the year of agents? Yes, but not the way the headlines suggested. Agents didn’t become tireless digital employees running your business while you slept. They left the demo stage and began working on real projects, and in the process we learned that handing something off is far harder than asking for it.
Agents won’t get better in isolation. We have to grow alongside them, and our tools and systems of record have to become legible to them. The winners of this era will be the companies whose systems know when we matter, when we’re needed, and when we’re finally free to step away.
