Six specifications decide whether an agent can move off the harness it was built on

The stickiness is in the loop. A harness holds the conversation, dispatches each tool call, checks it against a permission rule, manages context, and turns the model's text into work ^[1]. Every one of those steps writes state, and the state lands in different places: the system prompt on one screen, the tool grants on another, the context in a proprietary store, the approval rules in what the piece calls a settings page nobody remembers configuring ^[11]. That last location I believe without needing evidence.

What a specification does here is the trick Parquet did for files, Iceberg for tables and Polaris for catalogs: it turns a proprietary internal structure into a document any conforming system reads ^[10]. The trick has a boundary. The document carries the rule; the enforcement point stays inside the loop. Two harnesses can parse the same approval artifact and still disagree about which call trips it, because a spec constrains the file and not the runtime reading it. So the useful test for each of the six is a second implementation, preferably an unfriendly one, that reads the artifact and produces the same grants and the same refusals. And a spec is only half a migration path. The other half is the incumbent product choosing to emit the artifact, which no document can compel.

By that test the six are unevenly placed, and the source is explicit that they are not competitors, each answering a different question, with a complete system needing an answer to all six ^[12]. Three have outside stewards ^[5]. The other three, OAP, AGS and AAIS, were written by the author, who also works at Dremio, which ships an MCP Server as part of its platform ^[6]. He states this near the top and tells readers to weigh his enthusiasm accordingly ^[7], which is the right way to publish a spec you wrote. It still means half the coverage of a six-artifact stack rests on drafts with a single author ^[13], and the approval artifact is in that half.

Timing is what makes this expensive rather than merely irritating. The piece puts the wall at around month three ^[16] and dates real usefulness to the gap between day one and day 180 ^[8]. An agent becomes worth moving on the same curve that makes it hard to move.

Agent Skills is the one the source credits with adoption across directly competing vendors ^[15]. On the evidence supplied, that is the only artifact type here with any claim to having crossed between two runtimes whose owners have no interest in each other's success.