The Loop

The mechanism

What a loop actually is

A loop is the core unit of agentic AI. You give a model a goal, a set of tools, and a way to check its own work. Then you let it run. It takes an action, observes what happened, decides whether the result is good enough, and goes again. Act, observe, verify, repeat. It stops when the goal is met or when it needs a human.

This is different from the two things it replaces. A single prompt is one turn with no feedback: you ask, it answers, you check it yourself. A fixed pipeline is a set of steps you wired together in advance, with the model as one function call inside it. A loop hands the control to the model. Anthropic draws the line cleanly in its guidance on building agents: workflows follow predefined paths, while agents "dynamically direct their own processes and tool usage, maintaining control over how they accomplish tasks."

The part that makes a loop work, rather than wander, is verification. A loop with no way to check itself drifts. A loop that can test its own output improves with every pass. This is the single most important design choice, and the people building these systems are blunt about it.

The role shift

From doing the work to directing it

When the loop does the work, the human moves to the two ends of it. Upstream, you write the goal, the constraints, and the context the loop reads. Downstream, you decide whether to trust what came back. The middle, the production of the work itself, is no longer where a senior person spends their time. Osmani calls this the move from conductor to orchestrator: from "how do I solve this with the AI's help" to "what can I delegate, and how do I know it was done right."

This inverts where the bottleneck sits. For thirty years the scarce resource was the capacity to produce: to write the code, draft the report, build the model. Loops make production cheap and fast. What stays expensive is judgement. Deciding what to build, and deciding whether the result is safe to ship, are still human and still slow.

The data already shows the strain. As agentic coding spread through 2025 and 2026, raw output rose sharply while real productivity barely moved, because the work simply relocated. In one widely cited analysis, code output rose roughly fourfold while measured productivity gains stayed in the low double digits, and review time climbed by several hundred per cent. The work did not disappear. It moved from writing to checking.

For an enterprise, the lesson is direct. If you measure the value of AI by how much your teams now produce, you are measuring the cheap half. The scarce resource in your organisation is senior judgement, and a loop spends it faster than ever. Plan for that, or the bottleneck will find you.

The control plane

Verification is where control lives

Here is where a regulated firm has an advantage it may not realise. The verification gate, the point in the loop where output is checked before it is accepted, is exactly the kind of control financial services already understands. It is a checkpoint with a standard, an owner, and a record. You have built these for decades. The loop just makes them the centre of the operating model rather than an afterthought.

Tier the gate by blast radius, not by author. A loop that drafts an internal summary and a loop that moves money cannot pass through the same check. Score each loop by what it can affect if it is wrong, then set the gate accordingly. Low blast radius can clear an automated check. High blast radius needs a named human and a logged decision. This is risk tiering, which your firm already does for everything else. Apply it to loops.

The reason this matters is that a loop scales the consequences of a bad decision as fast as a good one. The same mechanism that lets one capable person direct a great deal of work will, with a weak gate, propagate an error across the whole portfolio before anyone looks.

The verification gate is your control plane. Designing it well is the most important piece of work in adopting loops, and it maps almost one to one onto the model risk, audit, and approval functions you already run. The deployment playbook covers how to wire monitoring, drift detection, and audit trails into that gate in practice.

The context layer

Your written-down judgement becomes the product

A loop is only as good as the context it reads. Engineers learned this fast: the highest-leverage file in an agentic codebase is the one that tells the agent how the team works, what to avoid, and what "good" looks like. Every time the agent makes a mistake, you do not correct it once. You write the correction into that file so it never makes the mistake again.

For a regulated firm, this is the most valuable artefact you will build. Your policies, your risk appetite, your controls, and your hard-won judgement about edge cases become machine-readable context that every loop reads before it acts. The institutional knowledge that currently lives in senior people's heads, and walks out of the door when they leave, gets written down in a form the work itself consumes.

There is a compliance dividend here. The context a loop read is itself a record. When the regulator asks why a system behaved as it did, "here are the policies it was operating under, versioned and dated" is a strong answer. Codifying your judgement is not only what makes the loop safe. It is what makes it auditable.

What to watch

The risks the loop introduces

Loops create new failure modes, and a serious firm names them before adopting them. The first is what Osmani calls comprehension debt: reviewing work you could no longer produce yourself. It is dangerously easy to approve output you do not fully understand. Over time a team can lose the ability to write what it still signs off, which is a new form of model risk and key-person risk at once.

The second is accountability. Under the Senior Managers regime, a person is responsible for outcomes. A loop cannot hold that responsibility. So the operating model has to keep a human on every load-bearing decision, with the authority and the understanding to refuse. Governance moves from a gate at the end of a project to a continuous part of how the loop runs, but the human checkpoints do not disappear. They become the point.

The third is cost and scale. Loops decouple throughput from headcount, which is the upside, but the same property removes a natural brake. One developer running agents in parallel reportedly spent over a million dollars on compute in a single month. Throughput is now bounded by compute and review capacity, not by how many people you employ. That has to be budgeted and governed deliberately, because nothing in the loop will stop on its own.

None of these is a reason to wait. They are the reason to design the loop, and the gate around it, with the same care you apply to any system that can move money or make a decision about a customer.

Where to start

Designing your first loop

Start small and start safe. Pick one process with a clear definition of done and a cheap way to check the result. A loop that reconciles two data sources, drafts a first-pass document, or triages a queue is a better first move than anything that touches a customer or a balance. You are not trying to prove the loop can do everything. You are trying to learn how to design the gate.

Front-load the specification, back-load the review, and let the loop run the middle. Spend real time making the goal and constraints precise, because that is the work that now pays back. Build the verification gate before you let the loop touch anything that matters. Then widen the blast radius slowly, as your confidence in the gate grows, not as your enthusiasm for the output grows.

This is the same discipline that governs any deployment in a regulated environment: one well-chosen process, done properly, before the next. The enterprise AI guide sets out how leadership should frame the decision, and the deployment playbook covers the operating model, architecture, and governance in detail. The loop does not change those fundamentals. It raises the stakes on getting the gate right.

The firms that win the next few years will not be the ones that produce the most. Production is cheap now. They will be the ones that built the best loops, and the best gates around them.