Operator running many agents unattended — ADOPT

Runs tens or hundreds of agent workers against shared infrastructure, mostly with nobody watching. Their pain is not any single run – it is that a percentage of runs quietly lie about having finished, and they find out days later. They will spend two engineer-weeks to make that stop. This is the shape fak was actually built for, and it shows.

• Next best option overall — a queue of agent workers against a shared inference endpoint, each trusting its own completion report
• Tolerance — 80 engineer-hours of unfamiliarity before this buyer walks away. Every cell’s load is measured against it.
• Switch bar — +0.50: the w_eff one axis must clear before switching is rational.
• Verdict — ADOPT. run-integrity clears the switch bar (+0.50)

facet	weight	w_eff	verdict	β	load	vs
Run integrity	0.25	+1.00	CRUISE	+1.00	+0.30	an agent loop that trusts its own completion report
Session longevity	0.15	+0.74	CRUISE	+0.76	+0.25	no context management – run until the window fills
Token economy	0.25	+0.72	CRUISE	+0.75	+0.25	tuned warm per-agent KV cache
Observability	0.10	+0.68	CRUISE	+0.67	+0.12	OpenTelemetry GenAI tracing (LangSmith, Langfuse, OTel)
Injection control	0.05	+0.55	CRUISE	+0.00	-0.50	formal-isolation defenses (CaMeL, MELON)
Raw speed	0.15	-0.51	DRAG	-0.25	+0.25	a raw SGLang endpoint, unmediated
Steady-state ops	0.05	-0.68	DRAG	-0.40	+0.25	a raw SGLang endpoint, unmediated

Run integrity — +1.00 (CRUISE)

When you are not watching, does the run end in a state you can trust and resume?

• fak 1 — MODELED, tools/loopgate verified-vs-naive-loop.report.json: the governed arm cuts false-done from 0.50 to 0.00 (net +3,900 tokens), with one-holder leases and runaway/injection caps; the ungoverned arm does none of it
• next best 0.5 — an agent loop that trusts its own completion report, the ungoverned arm of the same report: false-done 0.50
• ceiling 1 (definitional)
• β +1.000 → w +3.800
• adoption 24 h fak vs 0 h alternative, of 80 h → load +0.30 → τ +0.310
• w_net +3.491 → capped at CRUISE (MODELED evidence) → w_eff +0.999
• cost basis POLICY.md 1,726 w + docs/governed-agent-quickstart.md 1,361 w, plus wiring leases into an existing worker pool
• Read This is the reason this buyer exists as a segment. Half of unattended runs falsely reporting success is the operator’s actual daily problem, and it is the one axis where the mechanism reaches the definitional ceiling. It is also where the claim cap bites hardest: the raw arithmetic puts this cell near the top of the ladder and the MODELED cap pulls it back to CRUISE, because the corpus is authored. That gap between what the arithmetic says and what the evidence supports is the highest-value benchmark work in the repo.
• Fence The 0.50 baseline comes from an AUTHORED corpus in tools/loopgate, not from field measurement of a real fleet. The governing mechanism (leases, caps, external verification) is real and tested; the rate it is scored against is constructed. Capped at CRUISE for that reason – a simulated corpus cannot buy a category-defining claim.

Session longevity — +0.74 (CRUISE)

How long can one session keep going before context forces you to start over?

• fak 0.24 — OBSERVED, compaction_budget=48000 in docs/nightrun/cache-savings.jsonl – the resident context a session is held to – against a 200,000-token window
• next best 1 — no context management – run until the window fills, definitional: an unmanaged client holds the whole history, so it needs the whole window and then stops
• ceiling 0 (definitional)
• β +0.760 → w +0.996
• adoption 20 h fak vs 0 h alternative, of 80 h → load +0.25 → τ +0.255
• w_net +0.741 → capped at RELATIVISTIC (OBSERVED evidence) → w_eff +0.741
• cost basis compaction is on by default once the guard is running; no configuration beyond the serve setup
• Fence Scores the enforced resident budget, not work preserved across a compaction fire. For unattended runs the second is what matters and it is unmeasured.

Token economy — +0.72 (CRUISE)

For the same delivered work, how much of the input-token bill does it remove?

• fak 0.983 — MEASURED, BENCHMARK-AUTHORITY.md headline row, headline-qwen-50x5.json (50 turns x 5 agents): fak does 1 unit of prefill work where naive re-prefill does 60.3, eliminating 1 - 1/60.3 = 0.983
• next best 0.932 — tuned warm per-agent KV cache, the tuned arm of the same artifact: 4.1 units of 60.3, eliminating 0.932
• ceiling 1 (definitional)
• β +0.750 → w +0.973
• adoption 20 h fak vs 0 h alternative, of 80 h → load +0.25 → τ +0.255
• w_net +0.718
• cost basis docs/serve-config.md 3,286 w + AGENTS.md 10,166 w – the read to a configured fak serve fronting a shared endpoint; wiring an existing worker pool through the endpoint
• Read The 5-agent measurement matches this buyer’s shape exactly: the reuse being measured is CROSS-agent prefix sharing, which is the thing a fleet has and a single developer does not. This is the one segment where the headline number is quoted in its native regime.
• Fence Measured on Qwen2.5-1.5B Q8_0 on a single host. Radix prefix reuse is scale-free in principle; this ratio is not measured at frontier scale or against a production multi-tenant cache. The related RadixAttention row records an 86.7% hit rate, which is 100% of optimal on that corpus – the mechanism is at its ceiling, the workload is not the buyer’s.

Observability — +0.68 (CRUISE)

After the fact, from one artifact, can you reconstruct why it did what it did?

• fak 0.8 — OBSERVED, the session journal records the consequential decision classes – model traffic, cache reuse, compaction, tool-policy verdicts, recovery – post-hoc with no live attach. Scored 4 of 5, not 5 of 5: all five classes are RECORDED, but not in one artifact – the session journal, docs/nightrun/cache-savings.jsonl, and the /metrics split are separate surfaces, so a full reconstruction means joining across them. Everything is captured; single-artifact reconstruction is not.
• next best 0.4 — OpenTelemetry GenAI tracing (LangSmith, Langfuse, OTel), OTel GenAI conventions instrument the model call and partially the cache; no span exists for a compaction, a tool verdict, or a recovery decision. 2 of 5 classes
• ceiling 1 (definitional)
• β +0.667 → w +0.805
• adoption 20 h fak vs 10 h alternative, of 80 h → load +0.12 → τ +0.126
• w_net +0.679 → capped at RELATIVISTIC (OBSERVED evidence) → w_eff +0.679
• cost basis fak: journal on by default with the serve setup; alternative: a tracing stack across a fleet, which many operators already run
• Read For unattended fleets this compounds with run-integrity: knowing a run lied is only useful if you can then reconstruct where it started lying, and the tracing alternative carries no span for the verdict that would tell you.
• Fence Capability decomposition, not a benchmark; capped at RELATIVISTIC by the OBSERVED provenance rule.

Injection control — +0.55 (CRUISE)

Can hostile text inside the data the agent reads make it act for the attacker?

• fak 0 — MEASURED, BENCHMARK-AUTHORITY.md AgentDojo row: full-stack ASR 0/38 = 0.000 versus 29/38 = 0.763 detection-only
• next best 0 — formal-isolation defenses (CaMeL, MELON), CaMeL ASR 0, MELON 0.0-2.4% – already at the definitional floor
• ceiling 0 (definitional)
• β +0.000 → w +0.000 · mode parity_at_ceiling
• adoption 20 h fak vs 60 h alternative, of 80 h → load -0.50 → τ -0.549
• w_net +0.549
• cost basis fak: POLICY.md 1,726 w plus wiring existing workers behind the guard, no change to their data flow; alternative: restructuring every worker’s data flow so untrusted content cannot reach the control path, across a fleet
• Read Parity on the number; the score comes entirely from the fact that restructuring a whole fleet costs three times what wrapping it does.
• Fence Benign and under-attack utility on the same suite is NEEDS_KEY. ASR 0 bounds harm without showing the workers still do their job.

Raw speed — -0.51 (DRAG)

How fast do tokens actually come out, against the hardware’s physical limit?

• fak 0.75 — MEASURED, BENCHMARK-AUTHORITY.md gateway row, experiments/qwen36/gpu-server-r4-20260622/compare.json: fak-gateway 1085.6 tok/s vs raw-SGLang 1451.6 at C=64 = 0.75x
• next best 1 — a raw SGLang endpoint, unmediated, the raw-SGLang arm of the same comparison
• ceiling 1 (physical)
• β -0.250 → w -0.255 · mode pure_tax
• adoption 20 h fak vs 0 h alternative, of 80 h → load +0.25 → τ +0.255
• w_net -0.511
• cost basis same serve configuration as the token-economy cell
• Read Quoted at the worst measured point. This buyer in particular should read the convergence: the same row records the tax falling to ~3% at saturation (C=128, 0.97x), and a fleet operator runs saturated by definition. The conservative number is scored; the realistic one is better.

Steady-state ops — -0.68 (DRAG)

After you have adopted it, what does it cost every month to keep it running?

• fak 14 — MODELED, authored estimate from the measured adoption surface, 2026-08-10: 100 top-level verbs and 34 root operating documents layered on an existing fleet, with its own upgrades, lease semantics, and failure modes
• next best 10 — a raw SGLang endpoint, unmediated, authored estimate: a worker pool against a shared endpoint is already a substantial recurring bill
• ceiling 0 (definitional)
• β -0.400 → w -0.424
• adoption 20 h fak vs 0 h alternative, of 80 h → load +0.25 → τ +0.255
• w_net -0.679 → capped at CRUISE (MODELED evidence) → w_eff -0.679
• cost basis same serve configuration as the other cells in this segment
• Read The smallest ops penalty on the board in proportional terms, for the same reason this segment scores best overall: an operator already carrying a fleet absorbs the increment better than anyone else.
• Fence Authored estimates grounded in the measured surface. The ratio is the claim.