The AI Agent Productivity Gap: 220% More Code, 36% More Shipped

August 31, 2026. If a vendor is pitching your business AI automation on a headcount business case, two data sets published on the same day last week give you the question to ask back. Reuters reported on August 26 that Meta shelved the second wave of a plan to shrink some teams by up to 60 percent, after its own telemetry showed code changes to its AI platforms and infrastructure up 220 percent year over year while new or improved features that actually reached users rose just 36 percent. The same day, the infrastructure company Temporal published a survey of more than 550 engineers in which 80.8 percent said they now use AI agents daily or more. Adoption is real and it is accelerating. Delivered output is a separate number, and delivered output is the thing you are actually paying for.

What the two data sets showed

1. Meta ran the experiment at full scale and stopped it. Under the codename Project OT, short for Organization Transformation, Meta planned two waves of cuts, one in May and one in November, with some teams shrinking by as much as 60 percent and the remaining work supervised by small groups of staff overseeing AI agents. On the evening of May 19, hours before the first round, Mark Zuckerberg halted the November wave. Meta has said the deepest cuts were scenarios under consideration rather than a settled plan.
2. The activity number and the outcome number diverged by about six times. Code changes to Meta's AI software platforms and infrastructure rose 220 percent year over year. Features reaching users rose 36 percent.
3. The drag was measured too. Technical and security incidents rose 40 percent, and the time employees spent on those problems grew by 70 percent. That is the part of the bill that rarely appears in a pilot.
4. The chief executive said it out loud in July. At a company town hall Zuckerberg told staff the trajectory of agentic development over at least the previous four months had not accelerated the way the company expected.
5. Independently, engineers report the same friction. In Temporal's second annual State of Development report, fielded from April 29 to May 25 across more than 550 engineers and engineering leaders in the US, UK and EMEA, 41.1 percent said they hit agent related issues daily or more often, and 9.0 percent said continuously.
6. And yet confidence is very high. In the same survey, 91.1 percent said agents had improved or revolutionized their productivity and 85.5 percent trusted agent output at least somewhat. Daily use went from 47.3 percent a year earlier to 80.8 percent. The median respondent runs 5 agents, the average runs 10.7, and some reported well over 100.

The gap sits in the last mile, not in the model

Read the two sources together and they are not in conflict. They are measuring different layers. Temporal measured perceived productivity and activity: how many agents are running, how fast a prototype becomes production ready, how good it feels. On those measures the picture is genuinely strong, with 51.3 percent of engineers moving from AI prototype to production ready code in hours or faster and 26.9 percent in minutes or faster. Meta measured what arrived at the other end of the pipe: shipped features, and the incidents created along the way.

Where both data sets look at the same layer, they agree. Temporal's 41.1 percent daily issue rate is the survey version of Meta's 40 percent rise in incidents. Neither number is about model quality. Both are about review, integration, testing and cleanup, which is the part of the workflow almost nobody automates, and the part that absorbs the time the agent just saved.

Why this matters more to a 12 person business

Meta has close to unlimited budget, in house model researchers and the best available tooling. It still could not convert a 220 percent activity increase into a proportional output increase inside a year. A 12 person agency or a founder led e commerce brand has none of those advantages, so the honest planning assumption is that your conversion rate from agent activity to delivered work will be worse than Meta's, not better.

That is not an argument against automation. It is an argument against a particular business case. The business case that fails here is the one that says the agent replaces a person. The business case that survives is the one that says the agent shortens a cycle: lead response time, quote turnaround, ticket first touch, list build, reporting. Those are queue problems, and queues are exactly what agents are good at.

The build versus buy number nobody is quoting

One finding in the Temporal data deserves its own line on an agency's radar: 92.3 percent of engineers said they have tried to rebuild software they used to buy. That is not the same as succeeding, and the same survey shows why it often will not, but it changes the sales conversation. Your client's technical lead has probably already prototyped a replacement for something you charge for, and it probably works in a demo. The defensible position is the part the prototype does not cover, which is the operations layer this whole story is about: monitoring, incident handling, data quality and the person who is accountable when it breaks at 2am.

The same survey found engineers split on what this does to jobs. About 77.5 percent were more optimistic about their own role than a year earlier, yet 56.7 percent expected it to get harder for junior engineers to find work and 45.5 percent said the same for senior engineers, even though only 26.4 percent of companies reported slowing or stopping hiring. Sentiment is running well ahead of the actual hiring data, in both directions. If you are planning a team around this, plan around the hiring number.

What it means for operators

1. Ask for a shipped outcome metric, not an activity metric. Emails sent, drafts generated and tasks executed are the 220 percent number. Replies booked, tickets closed without a human reopening them and orders shipped are the 36 percent number. Make the proposal quote the second kind.
2. Price the incident tax before the pilot starts. Count reworks, escalations and hours spent correcting output for two weeks before you switch anything on, then count them again after. Meta's 70 percent rise in time spent on problems is the line item that quietly eats the saving.
3. Buy on cycle time, not headcount. Cycle time is measurable in a fortnight and it compounds. Headcount reduction is a claim you cannot verify until after you have made it irreversible.
4. Automate the queue, keep the judgement. Every deployment in the Temporal data that engineers trusted had a human decision point where the cost of being wrong was high. Put the approval where the money is.
5. Run a parallel control for 30 days. One team or one pipeline stays manual. Without a control you will attribute normal seasonality to the agent, in either direction.

This is how we scope work at imisofts. When a client asks for AI automation, the first deliverable is almost never an agent. It is the baseline: what the current cycle time is, where the queue backs up, and what a fix is worth. Agencies running the same play for their own clients can see how we structure it on our AI automation for agencies page, and teams that want the build done in house usually start by hiring an AI engineer for the instrumentation rather than the agent.

The counterargument, stated fairly

Three caveats belong on this. Meta's figures are internal telemetry described by Reuters and independently summarised by Engadget, not audited numbers, and Meta disputes the framing that it ever intended to remove 60 percent of its workforce. Temporal sells infrastructure for running agents reliably, so a finding that agents are hard to run reliably is convenient for it. And the Temporal survey closed on May 25, which is before several model releases that engineers now use daily, so the adoption figure is probably already low. The productivity gap may narrow. The point is that as of today it is measurable, and you should not buy against a number that nobody has shown you.

One more thing worth noticing. Meta's plan did not only fail on the technology. Staff who believed keystroke tracking software was training their replacements flooded internal channels with complaints and employee sentiment dropped by 19 points. If your automation plan requires your team to help build the thing they think will replace them, the technical risk is not the biggest risk you are carrying.