Spine Swarm makes the canvas the coordination primitive, and moves the hard part to the validator

Locks and validators sit on opposite sides of a gap the post names and does not close. Region locking with optimistic concurrency serialises writes: an agent asks for a lock, takes exclusive access if the region is free, otherwise waits or goes elsewhere, and stamps a version number on release ^[10]. That handles cell collisions. It does nothing for the case the post raises itself, where one agent moves a revenue growth rate while another moves a margin assumption, both writes succeed, and the model is now internally inconsistent ^[11]. Semantic consistency is handed to a validator agent that scans cross-cell dependencies while the run continues, rather than being enforced on the write path ^[9][11].

So correctness is eventual, and the interval is whatever the validator's scan loop happens to be. The supplied material puts no number on it.

Recovery has the same shape. The append-only log lets you replay the canvas to any point, but the post concedes it does not solve rollback, because reverting a bad artifact breaks the downstream artifacts that reference it, and says the system needs a repair workflow ^[7]. Undo is not on offer. Another agent pass is.

Two of the three handoff patterns require an agent to go and find its work: dependency detection has agents scanning the canvas for incomplete artifacts they can fulfil, and review loops have a supervisor reading finished ones ^[8]. The worked example shows the surface that gets scanned. Transcripts for five companies over four quarters is twenty artifacts written into regions ^[12][20] before the analyst has anything to summarise, with four agents and an orchestrator sharing one canvas and no agent calling another ^[12][14].

The observability claim is a spatial trace in place of a call stack, built from agent ID, timestamp, region coordinates, the cells and formulas that changed, and the reference edges between artifacts ^[5][6]. That tells you who wrote what, where, and when it landed. A call stack tells you why. When a chart comes out wrong, the log will name the chart agent and the table it read, and the reasoning that turned that table into that chart is not among the things stored ^[6].

Two caveats on the evidence. All of the above comes from a single post on dev.to under the mech_app_ai account ^[15], and the passage enumerating three observability layers breaks off partway through the first, so the other two are not described in what we have ^[13]. There is no lock granularity, no contention rate, no validator latency. For an architecture whose entire argument is that spatial coordination scales better than message passing for artifact-shaped output ^[2], those are the numbers that decide it.