Nishi multi-session coordination (ws_sync journal + claims + evidence-derived phase + debt-gate) vs AutoGen · CrewAI · LangGraph · Temporal · Linear/Jira · git — 12 axes · presence not depth · liar-killed. Live proof: /standup · /org. Snapshot 2026-07-19.
Across this working session, 5 concurrent agent sessions (Claude + siblings) drove ONE 180-task board and one ws_sync journal — zero clobbers. Sessions co-evolved shared organs (nx_ws_cycle, the debt plane) by claims + narrow pathspec, caught each other's regressions via the loop's RECON phase, and one session's banked finding became another's fix within the hour. The /standup board shows ACTIVE/STALE liveness derived from real beats.
| Axis | Nishi | SOTA (AutoGen/CrewAI/LangGraph/Temporal/Linear/git) | Verdict |
|---|---|---|---|
| Evidence-derived phase | phase computed from machine artifacts (gate hash, commit, edge bytes), never self-reported | all rely on self-reported status / task state | EXCEED |
| Concurrent multi-agent, zero-collision | N sessions, one board; claims + pathspec + flock leases; 0 clobbers proven | agent frameworks = one orchestrator; git = merge conflicts | EXCEED |
| Debt-gated work ordering | eat-debt-first mechanized: conductor REFUSES work while an in-scope debt ≥ floor is open | none gate new work on debt by construction | EXCEED |
| Work-liveness heartbeats | ACTIVE/STALE from real beats on /standup, human AND machine | Linear/Jira = manual status; frameworks = none | EXCEED |
| Agent-agnostic (vendor-neutral) | MCP is the contract; Claude, Kimi, role daemons interchangeable | AutoGen/CrewAI/LangGraph = single-framework | EXCEED |
| Durable / crash-safe | sovereign journal + seg-stores survive crashes; resume from artifacts | Temporal HAVE (external svc); most frameworks in-memory | HAVE |
| Provenance (who did what) | ws_sync actor + plane hist (epoch/actor/verb) | git commit author (but bot-sweep can mis-attribute) | HAVE |
| Shared memory / message-passing | ws_sync journal + planes = the shared substrate | frameworks HAVE (in-process); actors HAVE | HAVE |
| Autonomous (no human in loop) | sessions coordinate via journal + cadences, zero human | Linear/Slack human-driven; frameworks need a driver | HAVE |
| Claim/beat/release ON MCP (write) | READ live (nx_ws_cycle board/standup); WRITE via nx_fs_write edit (not a clean flock verb) | frameworks HAVE native APIs | GAP → F251b |
| Human-in-loop UI | /standup + /org read-only views; no interactive assign/comment | Linear/Jira = rich UI | GAP → F756 |
| Distributed / cross-machine claims | claims plane laptop-local; NAS-hub in flight (F251) | Temporal/Linear cloud-native multi-node | GAP → F251 |
Method: capability-presence census across 12 coordination axes; depth honestly discounted; gaps filed as owned rungs (F251 hub-claims, F251b claim/beat-on-MCP, F756 human-UI) per the less-than-SOTA-is-never-design law. Evidence: knowledge/status/ws_sync.jrnl (the live journal) + the /standup board + the frontier- / debt- planes. Every exceed is measured against what happened THIS session — not asserted.