Agentic AI and the future of work

AI · 2026-07-05 · 11 min read

By Arjun Mehta, Engineering & product

The jump from answering to acting is small in demos and enormous in production, because errors stop being wrong words and start being wrong actions.

An assistant that answers badly wastes a minute. An agent that acts badly sends the email, refunds the customer or drops the table. The capability gap between the two is modest; the consequence gap is not, and that is what makes deployment hard.

The demo hides this because a demo runs once, on a happy path, watched by the person who built it. Production runs thousands of times, on inputs nobody anticipated, while everyone is looking somewhere else. The interesting failures only show up at that scale.

Reliability compounds in the wrong direction. A step that succeeds 95 percent of the time is impressive in isolation and useless ten steps deep, where the chain completes barely six times in ten. Long autonomous runs need checkpoints, not better prompts.

The arithmetic is worth sitting with: 0.95 to the tenth power is about 0.60. Push the chain to twenty steps and you are below forty percent. This is why the agents that work in practice are short, or broken into stages a human or a hard check confirms before the next stage begins.

So design for reversibility. Let the agent read widely and write narrowly. Anything that leaves the building — a payment, a message, a deletion — passes a human or a hard rule first, and every action writes a log entry that can be replayed and undone.

A useful mental model is the blast radius. Reading a database, drafting a reply and summarising a document have a blast radius of nearly zero, so let the agent do them freely. Sending money or deleting records has a large one, so those get a gate. Sorting your actions by blast radius tells you where to spend your review budget.

The work that survives is specification and review. Describing a task precisely enough for a machine, then judging whether the result is actually right, is closer to editing than to typing, and it rewards people with deep domain knowledge rather than fast hands.

This is quietly good news for experienced workers. The skill that becomes scarce is not producing a first draft — the model does that — but knowing, at a glance, that the draft is subtly wrong. That judgment is expensive to acquire and hard to fake, which is exactly why it holds its value.

The teams getting real value picked one repetitive, well-bounded workflow with a cheap verification step, instrumented it properly, and expanded only after the numbers held for a quarter. The ones burning money started with the org chart.

If you want a place to begin, look for a task that is done many times a day, follows a stable recipe, and can be checked in seconds. Automate that, measure it honestly for a few weeks, and only then reach for the next one. Boring and bounded beats ambitious and unaccountable every time.

Tags: agents, workflow, automation

Arjun Mehta — Arjun writes about web performance, developer tooling and applied AI for ESPYCRUX. He has spent the last decade building and maintaining production web apps, and most of his articles start from something that broke in one of them.