← Understory

Why shared context beats competing summaries

When agents and humans work together, the expensive artifact is context — not output. Here's why we're building around that.

The common failure mode on teams working with AI agents isn’t that the agents produce bad work. It’s that every teammate ends up arguing from a different summary of what the agents actually did — each summary lossy in a different way — and the resulting debate is about the summaries, not the work.

Shared, full-fidelity context dissolves that failure mode. When everyone is reading from the same record of how the work actually happened, disagreement becomes productive again: we’re pointing at the same artifact and disagreeing about what it means, not disagreeing about what it was.

What we mean by full-fidelity

Not a transcript. Not a summary. A structured, queryable record of:

  • Every prompt, every tool call, every artifact the agents produced.
  • The decisions humans made between agent turns — and the reasoning attached.
  • The branches that didn’t ship, with enough of their context preserved that re-entering them later is cheap.

The thing you lose when context is lossy is the ability to disagree well. We’re trying to get that back.