Epoch had Opus 4.7 run unattended for 14 hours and it produced a software package estimated at 2-17 weeks of human engineering, for $251 of tokens. That is the price point that makes assigning work to an AI agentAn AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.Full definition → a budget line, not an experiment.
At OpenAI, a quarter of employees have four or more agents running at once in a given week, and legal, HR and other non-technical functions adopted agents at nearly the same rate as engineers.
Agents are the model plus a agent harnessThe scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.Full definition →: tool access, an environment to act in, and an app like Claude Code or Codex around it. A better harness raises the capability you actually get from a fixed model, so the wrapper is now part of the performance story.
Chinese open weightsA model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.Full definition → models sit on their own exponential curve, roughly 6-12 months behind the American closed frontier. Mollick warns their benchmark scores flatter them - real-world performance often trails what the numbers suggest.
On an exponential, each fixed window of change is bigger than the last. An AI plan written before winter 2025 assumed a couple of hours of work per prompt; months later it is sixteen. That is why a smooth curve is experienced as a series of shocks.
Terms in this piece · Glossary
grounding — Tying a model's answers to checkable sources — retrieved documents, live data, tool results — instead of letting it answer from memory alone.
AI agent — An AI system that doesn't just answer once but works toward a goal in a loop — taking actions, reading the results, and deciding what to do next.
agent harness — The scaffolding around a model that turns it into a working agent — the loop, the tools it can call, and the rules for when to stop.
open weights — A model whose trained parameters are published for anyone to download and run — unlike API-only models you can access but never possess.
Why it matters
Mollick's argument isn't another 'prompt better' guide — it's a data-groundingTying a model's answers to checkable sources — retrieved documents, live data, tool results — instead of letting it answer from memory alone.Full definition → case that the chatbot-as-interface era is ending, with capability gains accelerating even as the frontier stays jagged and unreliable in specific spots. Read it if you're deciding how much to bet your workflow on current tools versus what's landing next quarter; you'll come away with a sharper mental model of where models are actually strong, where they still break, and how fast that line is moving.
Key quotes
“Epoch , recently found Opus 4.7, working on its own for 14 hours, was able to build a software package that would take 2-17 weeks of human engineering work (it cost $251 in tokens).”
“We are moving from a world where non-experts use chatbots to fill in gaps to one in which experts use agents to get work done.”
“We are very bad at feeling exponentials from the inside, and we are currently inside one.”
“A quarter of OpenAI workers have at least four agents running at one time every week.”
