Three named agents shipped in 27 days, and none of them can read each other's config

The interesting part of a portable agent profile is not the copying. It is the part that refuses to copy. Take the simplest durable role anyone actually wants, a reviewer that reads code and does not write it. In the account given by the OAP author, that constraint shows up as a checkbox in one tool's interface, a flag in a second, and a paragraph of natural-language instruction in a third, and while all three ask the model to respect it, only some of the three enforce it ^[7]. A file that moves between those harnesses moves the sentence, not the guarantee. Whichever harness in your roster treats permissions as advice is the one that defines what the role can really do.

The window is tighter than the essay's own framing. It calls the collapse of the session model a five-week affair; the dated releases run from July 21 to August 17, which is 27 days ^[12]. The counting is loose in the other direction too. The same piece describes setting the reviewer up in three named tools and then calls the result a configuration scattered across five ^[6], which reads as a working estimate rather than an audit. Take the estimate at its own terms and it still gets worse with every tool added: reaching parity for one role across the fourteen agent CLIs the author reports installed would mean writing that role fourteen times ^[15].

Session-shaped tooling absorbed that waste because the context was cheap to rebuild. It stops absorbing it when a task runs four hours across 200 tool calls with nobody watching each turn, and when several agents hand work between themselves ^[9]. A handoff only works if each agent knows who it is, what it is permitted to do, and what the others have already learned ^[10], and under the old arrangement that last part was a per-tool cache: notes in a hidden directory here, compacted context there, and in some harnesses nothing at all ^[8]. Persistence turns that cache into an asset with no export.

Which is where the argument needs its discount applied. The person diagnosing the missing format is Head of Developer Relations at Dremio and the author of the Open Agent Profile specification and the tools that implement it ^[11]. He says so plainly, and the diagnosis holds up on the evidence he lays out. But the source records no sign that Block, xAI or Nous intends to read anyone's profile file, including his ^[14]. He also notes the three releases agree on the destination and disagree on nearly everything about the route ^[13], and a format only removes lock-in if the vendors doing the locking implement it. Absent that, the file is a convention for the tools that want to interoperate, and the harnesses holding four hours of accumulated context are not obviously in that group.