Writing

What Changes When AI Agents Close the Loop

Why AI agents that close the loop change how work is designed, and why orchestration is the hard engineering problem.

AI · 5 minute read

The wrong kind of AI

Most companies building AI are building assistants: a chatbot that drafts emails, a copilot that suggests code, a dashboard that surfaces insights you still have to act on. These products make human effort quicker, but they leave in place the structure that requires the effort. I think of that as a faster typewriter, and I would not call it a transformation.

The distinction matters more than many executives realize. An AI tool augments a person, which is a productivity gain. An AI agent takes over the person's place in a workflow: it perceives context, makes decisions, chains actions across systems, and closes the loop without waiting for someone to click approve. That means redesigning how the work gets done.

I made this argument on a panel about agentic AI at the University of Central Punjab in Lahore. Companies investing only in assistants are optimizing for the way work was organized a few years ago, while companies investing in agentic architecture are building for the way it will be organized a few years from now. The two groups can look like competitors in the same race, but they are solving different problems.

What agentic actually means

Agentic AI is a way of designing systems more than a product category. An agent can perceive its environment, form a goal, select and execute tools, evaluate the results, and iterate, without a person in the loop for each step. An assistant hands control back to the person after every step, and an agent keeps going until the job is done.

Being first to deploy a chatbot will matter far less than learning to coordinate agents that direct other agents.

Consider what this means structurally. In a traditional SaaS workflow, a person reads a report, decides what to do, opens a tool, and does it. In an agentic workflow, an agent watches the data, reasons about context, calls APIs across five systems, executes, logs the outcome, and notifies a person only when there is an exception. The person is no longer on the critical path, and their attention goes to the cases that need judgment.

The primitives that make this possible have matured quickly: function calling, persistent memory, multi-step reasoning, tool orchestration, and communication between agents. It is tempting to treat these as features to add to an existing product. I think they are better understood as load-bearing components of a new software architecture. Combined correctly, they produce software that uses other software, autonomously and continuously, where before we built software for people to use.

The architect shift

The uncomfortable part is that the human role in many knowledge workflows is becoming architectural where it used to be operational. Nobody configures a bridge every morning. You design it once, load-test it, and let it carry traffic. Agentic AI does something similar to white-collar work. The person who used to run the process becomes the person who defines the rules, constraints, and escalation logic of the agent that runs it.

Take candidate screening as an illustration. An agentic workflow can ingest CVs, score them against the criteria for the role, check them against public professional profiles, draft the reasoning behind a shortlist, and schedule interviews, with a person stepping in at the hiring manager's final review. What has been automated there is more than a list of the recruiter's tasks. It is the recruiter's place in that workflow. The recruiter's work moves to setting the criteria and deciding what should be escalated, which was always the part that needed the most judgment.

This is a hard shift for most organizations because it asks for a different kind of thinking. An operator asks how to do the task faster. An architect asks what the rules, exceptions, and failure modes are for the system that does the task. Most companies have not yet trained their people, or their leadership, to think about AI this way. Many are still hiring for prompt writing when the scarcer skill is systems design.

The organizations that make this shift early will be hard to catch. The models themselves are becoming a commodity quickly, so access to a better model is unlikely to be what separates one company from another for long. Orchestration logic is different: it encodes how a particular business works, and it is slow to copy. Every company will have access to a capable model. How it is wired into the business is where the difference will show.

Why orchestration is the hard part

There is a great deal of talk about agents and much less serious talk about orchestration. A single agent completing a single task is a demo. Multiple agents reasoning over shared context, handing off state, recovering from failures, and coordinating toward a business outcome make up a system, and systems have emergent failure modes that demos never show.

Orchestration is where I expect many enterprise AI efforts to break. The questions are unglamorous, and they are foundational:

  • How do agents share state without corrupting it?
  • How do you define trust boundaries between agents with different permission levels?
  • How do you build retry logic that accounts for non-deterministic model outputs?
  • How do you audit what a fully autonomous system decided, and why?

None of these is a model problem. They are software engineering problems, and many AI teams are not yet staffed to solve them.

I expect the companies that work out agent orchestration, and go beyond simply deploying agents, to pull ahead in their markets. The gain comes less from better answers than from better throughput. An orchestrated system keeps running between meetings and overnight, and each cycle leaves logs and corrections that improve the next one. A competitor who starts later has to make up every one of those cycles.

More writing

  1. Scaling Beyond the Founder: Systems for GrowthA company has scaled only when it makes good decisions without its founder in the room, and that takes systems more than headcount.4 minute read
  2. Choosing What to IgnoreClarity is scarcer than capital or talent, and it comes from choosing your inputs, and your rooms, on purpose.5 minute read
  3. Why Constraint Produces Better Architecture Than AbundanceBuilding software under real limits, as we do in Lahore, forces earlier and better design decisions than abundant capital does.4 minute read

If you have a reply to this, or a correction, I would like to read it. Write to me.