Why AI Agents Fall Apart After Hours of Work

Most AI agents can run for hours, but almost none can tell you if the work they did was right. Learn how Nuno Campos is solving this problem at Witan Labs.

Aug 19, 2026 32:04

Nuno Campos

Expert Insights

Nuno Campos spent 3 years at the center of the agent-tooling boom. He was the first employee at LangChain, where he created LangGraph: the tool that helps AI agents run for hours without falling apart. Now, as Co-Founder and CTO of Witan Labs, he's rebuilding Excel and PowerPoint from scratch, so that agents can work through files and verify their output on their own.

Hear him outline:

Why running an agent for 12 hours is easy, but getting it to produce something useful at the end is the hard part
Why the gap between a co-pilot and an autonomous agent comes down to runtime
What "oracle-driven development" means, and why Microsoft Excel itself became Witan Labs' ground truth
How a 7-person team divides coding-agent work using Slack-shared plan files and a self-updating internal handbook
Why sidebar-style plugins like Claude for Excel can't run unattended, and how a headless alternative helps
Where the model breaks down in practice: real use cases in finance, from nightly reconciliation to investment-bank modeling

Technically making an agent run for 12 hours is not hard. What's hard is making it run for 12 hours and producing something useful at the end.

— Nuno Campos, Co-Founder and CTO, Witan Labs

Monterail Team Analysis

Nuno's "runtime" framing gives teams a way to decide when a project needs an autonomous agent: ask how long it can run before someone needs to check in.
Oracle-driven development works beyond spreadsheets. Any domain with an authoritative, programmatically controllable system of record could become the "oracle" an agent tests itself against.
Designing verification loops is quickly becoming as important as writing the code itself. This is something agentic engineering teams, including our own, are increasingly expected to build alongside product features.
Witan Labs is betting against sidebar plugins. Worth watching if your team is evaluating copilot-style AI tools. Running server-side means it can work overnight or in batches, without a human at the keyboard.
The "editable handbook" pattern (agents documenting their own discoveries for future runs) lets engineering teams build up institutional knowledge without asking humans to maintain it.
Financial services (trade reconciliation, investment-bank modeling) stand out as an early, concrete vertical for headless document agents. Worth watching as a proof point for adjacent regulated industries.