Agentic AI in 2026: what changes when the prompt becomes a workflow

Every trend list for 2026 has the same item at the top. Gartner files it under "autonomy and agency" and names multiagent systems as one of its strategic technology trends for the year, with a forecast that by 2028 most B2B buying will run through AI agents. The narrative is that the era of the one-off prompt is over and AI now orchestrates work end to end with a person checking in rather than typing along. That narrative is roughly right, and the word "roughly" is where the engineering lives. This post is about what the shift actually requires in code, when more than one agent is worth the trouble, and what to build first.

The prompt did not go away, it got a loop around it

The difference between a chat feature and an agentic one is not intelligence. It is the loop. An agent calls a tool, reads the result, decides what to do next, and repeats until the job is done or it gives up. We laid out that definition in what are AI agents, and it still holds, because it predicts everything that follows. Once a model is in a loop with real tools, three things that did not matter for a chat box become the whole job.

State. A workflow that takes forty steps needs to remember what it has done, what it learned, and what it is still waiting on. Prompts do not remember. Something has to, and that something is a store you design, version, and can inspect when a run goes wrong.

Permissions. A chat model that writes a bad answer wastes a minute. An agent with write access to the CRM that misreads a record does damage that has to be undone. Every tool the agent can call is a decision about what it is allowed to touch, and the scope should be the narrowest that still lets it finish.

Stopping. The loop needs a budget in steps, tokens, and money, a timeout, and a rule for when to hand the task to a person. Without those, an agent that is confused does not fail; it keeps going, and the bill keeps going with it.

Together those three are the harness, and the harness is the part you own. The model is rented and gets replaced every few months. Our agent harness post covers what goes in one.

What "end to end" means in practice

The workflows that have crossed from demo to production this year share a shape. They start from an event, not a chat message: a document arrives, a ticket is opened, a nightly job runs. They touch several systems through tools with narrow permissions. They produce an artefact a person can inspect, such as a drafted reply, a filled form, a pull request, a reconciled ledger entry. And they end with a decision that is either automatic because the confidence is high and the action is reversible, or routed to a person because it is not.

Invoice intake is the canonical example. Read the document, extract the fields, match them to a purchase order, flag anything that does not reconcile, post the ones that do. Every step is a tool call, every tool has a scope, and the only step that needs a person is the exception queue. That is agentic AI. It is also, if you squint, ordinary workflow automation with a model doing the parts that used to need a human reader. The label is new. The engineering discipline is the same one that made any integration reliable.

When multi-agent is worth it, and when it is a slow way to fail

The trend lists talk about collaborative multi-agent systems as though a team of agents were the default. It is not. Most tasks are best served by one agent with good tools, a clear plan, and a harness, for the reasons in one agent is usually enough: every hand-off between agents is a place for context to be lost and for two models to disagree with each other.

Multiple agents earn their complexity in three cases. The work genuinely decomposes into parallel pieces that do not need each other's context, such as researching ten suppliers at once. The pieces need different permissions, so a read-only research agent hands a summary to a narrowly scoped acting agent rather than one agent holding every key. Or the pieces need different models, with a small cheap model doing triage and the expensive one doing the hard cases, which is routing wearing an agent costume. If none of those apply, one agent with a to-do list will beat a committee.

The number that decides whether it ships

Agentic features are judged on task completion rate, not on how impressive a single run looks. A workflow that finishes 70 percent of tasks correctly and routes the rest to a person cleanly is a product. One that finishes 90 percent and silently gets the other 10 wrong is a liability. You cannot know which one you have without an evaluation set built from real cases, scored before the feature ships and re-scored every time the model or the prompt changes. Evals before features is how we do it, and it is the first week of any agent build we take on.

The second number is cost per completed task. Agents burn tokens in loops, and a run that retries five times costs five times as much. Budget per run, cache the stable context, and route the easy steps to small models. Then measure what a finished task costs and compare it to what the person it replaces costs. That comparison is the business case, and it is usually good, but only once the retries are under control.

What to build first

1. One workflow, not a platform. Pick the process with the most repetitive reading and the clearest definition of done.
2. The evaluation set, from fifty real cases, before any prompt is written.
3. The harness: tools with narrow scope, a state store, budgets, and a human hand-off path.
4. Observability from day one, with every tool call and its inputs logged, as in what to log. When an agent does something odd, and it will, you need the trace.
5. Ship to a small group with a person reviewing the exception queue, then widen.

Agentic AI is the defining shift of the year in the sense that the systems people want are now loops with tools rather than boxes with prompts. The teams that get value from it are the ones treating that loop as software to be engineered, not as a model to be prompted. If you want help scoping the first workflow, that is a conversation we have most weeks.