# Evolution 85 — Agent Workforce Architecture

**FIRME AI CORPORATION · CAIN-42 · Evolution 85** — branch `evolution-85`, baseline Evolution 84 (`2c419a042`). Development result. Not deployed. Not released. Built by the three-person team.

## Three numbers that are never merged

| Quantity | What it is | Largest value measured |
|---|---|---|
| Registered identities | durable Ed25519 records | 10,000 |
| Active worker processes | real OS processes holding leases | 500 |
| Concurrent model inference requests | model calls in flight | 1 (one model resident, CPU) |
| Distributed workers (other hosts) | workers on a second machine | 0 in the stages; one second-host resource process in the governed demo |

## Identities (records, not workers)

| Stage | Registered | Seconds | Store KB |
|---|---|---|---|
| 10 | 10 | 0.003 | 2 |
| 50 | 50 | 0.002 | 13 |
| 100 | 100 | 0.003 | 25 |
| 250 | 250 | 0.005 | 64 |
| 500 | 500 | 0.009 | 129 |
| 1,000 | 1,000 | 0.018 | 258 |
| 2,500 | 2,500 | 0.045 | 646 |
| 5,000 | 5,000 | 0.101 | 1293 |
| 10,000 | 10,000 | 0.454 | 2586 |

## Real worker processes with a GLOBAL kill at 50% of the work

Latencies in ms. "Kill latency" is from the end of the engage transaction to each worker observing it; "from request"
adds the engage transaction itself. "Commits after engage" are tasks whose worker passed its pre-commit check before
the engage became visible: the measured in-flight window.

| Workers | Done/total | Tasks/s before kill | Task p50/p95/p99 | Queue p50 | Kill p50/p95/max | Max from request | Commits after engage | Aborted in flight | Duplicate effects | Done by revoked agents | PSS MB | CPU s | Storage KB |
|---|---|---|---|---|---|---|---|---|---|---|---|---|---|
| 10 | 170/300 | 83.47 | 68.432 / 150.157 / 184.925 | 969.5 | 63.47 / 79.94 / 84.42 | 87.87 | 0 (0.0 ms) | 8 | 0 | 0 | 91.6 | 6.48 | 4402 |
| 50 | 1050/1500 | 71.07 | 347.662 / 1038.583 / 1857.224 | 5267.4 | 182.28 / 376.26 / 405.63 | 407.04 | 5 (92.95 ms) | 43 | 0 | 0 | 221.6 | 40.55 | 5065 |
| 100 | 1640/3000 | 75.97 | 264.393 / 3823.418 / 6548.835 | 10906.2 | 163.25 / 371.82 / 433.38 | 440.83 | 71 (433.21 ms) | 27 | 0 | 0 | 404.1 | 64.67 | 5710 |
| 250 | 3890/7500 | 65.07 | 1015.856 / 9860.207 / 15998.903 | 31478.5 | 528.89 / 1263.48 / 1453.52 | 1455.44 | 192 (1452.39 ms) | 54 | 0 | 0 | 1136.4 | 176.61 | 7695 |
| 500 | 7646/15000 | 42.49 | 4372.362 / 21795.477 / 28101.144 | 94150.0 | 2556.77 / 4603.34 / 5372.41 | 5373.96 | 328 (5215.49 ms) | 123 | 0 | 0 | 2608.9 | 444.38 | 11283 |


{'1000, 2500, 5000 worker processes': 'not run on this shared production host; 10,000 is measured as registered identities only'}

## Endurance (actual wall clock)

| Item | Value |
|---|---|
| Duration | 3600.0 s actual (requested 3600) |
| Worker processes | 8 at a time, 24 started over the run |
| Tasks | 14,400 done of 14,400; 0 dead letters |
| Worker crashes injected (SIGKILL mid-lease) | 11 |
| Agent revocations mid-run | 5, all stopped: True |
| Duplicate effects | 0 (effects = done: True) |
| Governed receipts issued during the run | 120, all effect-verified: True |
| Checkpoints | 59, chain intact: True |
| Final GLOBAL kill | workers alive afterwards: 0; next action: DENIED |

6 h, 24 h and 72 h were not run. The endurance tasks are deterministic; the run makes no model calls and is a
low-duty-cycle test of recovery and revocation, not of intelligence.

## Mechanisms reused

Leases, fencing tokens, idempotent effect table, retries, dead letters and hash-chained checkpoints are the E81
`MissionStore`. E85 adds the kill check before a claim and again before a commit.

## Not built

Hierarchical swarm planning, capability-based dynamic assignment, cross-swarm conflict resolution and a multi-host
worker pool. The agent categories in the brief (commanders, planners, curators and so on) exist as roles in identity
manifests, not as evaluated behaviours.
