Your AI Agents Aren't Faster. They're Just Louder.

There is a chart going around right now that everyone in your group chat has shared at least once. It is from a small evaluation lab called METR, and it shows the time horizon of tasks AI models can complete doubling roughly every seven months. The line goes up and to the right. It looks inevitable.

On April 25, 2026, METR's president Chris Painter and technical staff member Joel Becker went on Bloomberg's Odd Lots to walk through the chart and what it actually measures. It is a great episode. It is also, accidentally, the most damning case yet that the AI industry is selling executives a story their own engineers cannot reproduce.

Here is the part nobody is reposting. METR also ran a randomized controlled trial with 16 senior open-source developers. Participants believed AI tools made them roughly 20% faster. When measured, they were 19% slower. That is a 39-point gap between perception and reality, on the population most likely to be sophisticated users of these tools.

Your AI agents are not faster. They are louder. And until you have real AI agent performance tracking in place, you cannot tell the difference.

The vibes economy has a measurement problem

The dominant unit of evidence for "AI is working" inside most companies is a Slack screenshot. Someone shipped something cool, posted a clip, and the rest of the org reasoned upward from there. Quarterly business reviews now feature productivity claims that are, charitably, vibes laundered through a bar chart.

This is not a Tommbo opinion. This is what the METR study revealed in the most uncomfortable way possible. Senior developers, working on real codebases they knew intimately, were systematically wrong about whether AI made them faster. They were not lying. They were not stupid. They simply had no instrumentation. They had vibes.

The seven-month doubling chart Painter and Becker discussed measures something real, but it measures it in METR's harness, on METR's curated tasks, with METR's prompts. None of that is your harness. None of that is your task distribution. None of that is your retry logic, your tool budget, your memory backend, or your timeouts. You cannot infer your stack's performance from a benchmark you did not run.

Which is the part that should make every leader uncomfortable: most companies that have spent the last 18 months deploying agents have less data on those agents' real behavior than METR has on a public benchmark. The viral chart is louder than your own product.

What real AI agent evaluation looks like (and why most teams are skipping it)

The reason METR's chart goes viral and your internal eval does not is that METR's chart exists. Most internal evals do not. When we talk to teams shipping agents into production, the typical state of AI agent evaluation looks like this:

• A spreadsheet of "good" and "bad" prompts that someone curated three months ago.
• A single accuracy number, usually pulled from a vendor's marketing page.
• A "vibes pass" before each release, where two engineers manually try the agent on six tasks and call it good.

That is not evaluation. That is theater. Real AI agent evaluation has properties most teams have not yet built for:

• It is task-distribution-aware. You measure on the work your users actually do, not the work the model provider's eval suite tests for.
• It is multi-step. You instrument every tool call, every retry, every hand-off, and you score the trajectory, not just the final string. An agent that "got the right answer" via four hallucinated tool calls is a liability, not a win.
• It is calibrated against humans doing the same task, in your environment. This is the part that hurt METR's study participants. You cannot grade an agent on whether it succeeded. You have to grade it on whether it succeeded faster, cheaper, or more reliably than the human path it replaced.
• It runs continuously. Models change under you. Your prompts drift. Your tools change versions. A one-shot eval at launch is a photograph of a stock price.

If you do not have this, what you have is an agent that occasionally produces an output your team likes, accompanied by a story your CFO is about to start asking hard questions about.

The compounding math everyone keeps forgetting

There is a simpler reason the gap between "AI agent benchmarks" and "AI agents in your business" is so large, and it is just arithmetic. Multi-step workflows compound errors.

An agent with 95% per-step accuracy, executing a 10-step workflow, succeeds end-to-end about 60% of the time. Drop per-step accuracy to 85%, which still sounds great in a deck, and end-to-end success collapses to roughly 20%. You do not feel this in a demo, because demos are short. You feel it in week six of production, when your support queue is full of "the agent did six things right and then deleted a customer record."

This is why AI agent monitoring has to live at the trajectory level, not the response level. The interesting failures are not "the model said something wrong." The interesting failures are "the model said something plausible, then chained that into three tool calls, and the third one was unrecoverable." A logging pipeline that only captures final outputs cannot see this happening. A dashboard that only reports end-to-end success cannot tell you which step is dragging the workflow down.

The METR research is, in a sense, an industrial-scale version of what every team should be doing internally and is not. They built a harness. They ran a controlled study. They measured perception against reality. They published the gap. The reason their chart is "viral" is that it is one of the only places that data exists.

The next 12 months belong to the teams who actually measured

Here is the contrarian read on the METR moment. Everyone is treating the seven-month doubling line as the headline. The real headline is the 39-point perception gap.

That gap is the most actionable number in AI right now, because it tells you something specific about your competition. Half of your peers are running AI initiatives based on perceived speedup. The other half do not even know whether their agents are working. The teams that win the next 12 months are not the ones with the best models. The model layer is commoditizing in real time. The teams that win are the ones who built AI agent analytics that can answer three questions for a CFO without flinching:

• Which workflows are actually faster, by how much, and on whose tasks?
• Where in those workflows are we losing reliability, and is the trend getting better or worse?
• When a model provider ships a new version next Tuesday, will we know within a day whether our agents got better, worse, or differently broken?

Most teams cannot answer any of these. The teams that can are quietly compounding an advantage that, like METR's chart, will look obvious in retrospect and impossible to catch up to in real time.

This is the bet Tommbo is built around. AI agent monitoring, evaluation, and performance tracking are not a feature you slap on once you have shipped. They are the system that decides whether the thing you shipped is real. The viral chart is fun. The doubling curve is interesting. Neither of them tells you whether your agent is doing its job today. Your instrumentation does. If you do not own that instrumentation, you are running an AI strategy on someone else's benchmark.

The METR team did the work and got a chart out of it. You can do the same work, on your own stack, and get a business out of it. But not if you keep mistaking volume for velocity.

Stop measuring how loud your agents are. Start measuring whether they're actually faster.

Source: Understanding the Most Viral Chart in Artificial Intelligence - Odd Lots (Bloomberg), April 25, 2026.