clm-0008

Source: a white paper describing 8 autonomous OpenClaw sessions ("Council of Minions") self-organising into 3 teams over ~3 hours, coordinated via a shared MongoDB brain and sessions_send, at ~$0.10 of tokens. THE INSIGHT WORTH KEEPING: "an orchestrator coordinates tasks; a council coordinates attention." Their strongest evidence is that nobody had been ASSIGNED to notice auth was added to a new engine and never backfilled to the old one — a task decomposition can only cover work you can name in advance. That limitation is real and matches our own experience: the R1 scorer artefact was caught by reading transcripts, not by any metric we had designed. SCEPTICISM: the headline ROI ($0.10 vs EUR 1,500-4,000 of consultancy) prices output volume as professional deliverable. "79 of 98 routes unauthenticated" is a grep. The cross-verification is agents agreeing with agents, with no independent validation. And the emergence is partly pre-seeded — an identity file that says "Auth, Secrets and Identity Security" makes a credential audit declared rather than emergent. Their gateway was also at ~50% RAM across 555 sessions, which is not a healthy system. WHY WE CANNOT DO IT TODAY, AND IT IS NOT ABOUT POWER: their architecture trades model capability for concurrency; ours does the opposite. MTP forces -np 1, so we have ONE KV slot behind one very large model. Eight concurrent sessions would each cold-prefill and destroy the warm lane. The prerequisite is many CHEAP WARM LANES, not more compute. THE OPERATOR'S EXTENSION (2026-08-05), which is the stronger form: use several DIFFERENT small specialist models rather than many instances of one. The paper ran 8 copies of a single model, so all 8 shared its blind spots — diversity of attention, not of capability. Different models have genuinely different failure modes, which is the same principle underpinning our blind-judging methodology. Our own field showed real profile differences: gpt-oss-120b terse to the point of under-delivery, 35B-A3B best on conceptual quality, 27B strongest on code. FEASIBLE SHAPE once aihydra lands: 122B deep reasoner (aibeast) + 35B-A3B fast generalist and a coder model (aihydra) + a small classifier on the M4 mini. Four genuinely different skills across three machines — heterogeneous by hardware as well as by prompt. ⚑ CORRECTED 2026-08-05 (operator catch). I filed this as a design conflict between self-cycling and Warden's restraint. That was the wrong layer. Restraint lives at the DELIVERY GATE (loop3: daily_cap, quiet hours, validity check, teaser-first), not in deliberation. Deliberation is already separated from delivery by design — lots may happen in the background; the operator is involved only when something of value needs delivering AND it is the right moment. So a self-cycling council in loop2 costs COMPUTE, not attention, and the conflict dissolves. THE RESULTING SHAPE — self-cycling as the first escalation tier: sentinel (cheap, continuous, high-recall, deliberately low-precision) flags a change -> council self-cycles on that one thing, bounded -> exits, and only then does loop3 decide if and when the operator hears about it. THREE EXITS, not two. The operator named "nothing is wrong" and "value delivered"; the third is BUDGET EXHAUSTED, INCONCLUSIVE — and it is the one that bites, because both clean exits are conclusions and real investigations often reach neither. Without it a council does not fail loudly, it quietly eats the box. BOUNDARY PRECISION: the council does not deliver. It produces a CANDIDATE WORTH QUEUEING. Loop3 remains the only path to the operator. A council that believes it can deliver has escaped the gate. FAILURE MODES PER TIER: sentinel too sensitive -> constant convening, cost; too conservative -> the 6h blind spot returns; council unbounded -> runs forever; council concluding "nothing wrong" too eagerly -> misses the thing. OUR OWN INCIDENT IS THE SPECIFICATION: inc-0004 saw the box crash-looping for 46 minutes with three llama-server kills before the kernel panicked, and nothing surfaced any of it — with the risk flagged in that morning's brief. Three process kills in 46 minutes is exactly what a change-detector flags, and "why does this keep restarting" is exactly a bounded investigation with a clear conclusive exit. TRANSFERABLE ENGINEERING (the most useful part): their inter-session sends timed out at a ~10s gateway default against 30-60s cold starts, and the fix was retry-with-backoff plus spill-to-disk staging rather than a larger timeout. That is directly applicable to our own scheduler-preemption work.