The autonomy spectrum
“Autonomous” is not binary. Most real systems sit somewhere on a spectrum defined by how often a human has to intervene:
The jump that creates business value is Steered → Autonomous — and the thing
that blocks it is almost never raw model capability. It’s the context layer and the
harness around the model. See Human steering.
Removing the human in the loop
“Remove the human” is a misleading goal if it means no human ever. The right goal is to move the human from the inner loop to the outer loop — from approving every step to reviewing outcomes and correcting the system. Three things have to be true before that’s safe:1
The agent can find the right context on its own
No pre-injection, no human deciding what to paste in. The agent searches and
discovers. See Context search accuracy.
2
Its context is trustworthy and traceable
Facts are current, non-conflicting, and carry provenance, so an outcome can be
audited without a human reconstructing what the agent “must have” used. See
Provenance and Conflicts & staleness.
3
The harness catches its own failures
Soft gates stop early exits; critic sub-agents catch bad reasoning in-session,
before it reaches the user. See The harness.
Domain agents raise the bar
General agents can lean on the model’s world knowledge. Domain agents — legal, healthcare, sales, infrastructure — can’t: the knowledge they need is private, structured, and changing. That’s exactly why the context layer matters more the more specialized the agent is, and why a structured context engine tends to be the highest-leverage investment for a domain agent.Next: The harness
What sits around the model to make the loop reliable.