Real run · 5 August 2026 · one Mac · zero cloud calls

The $0.00 Swarm

Six agent runs on the free local brain — four workers in parallel, a dispatcher to merge, and a reviewer to check the lot. They audited a real Agent OS: every number below came off the disk, and the meter never moved. This is the part a metered API quietly talks you out of: running a whole team at once.

0build folders audited
0gallery assets catalogued
0AI employees counted
0guide pages indexed

The parallel run — real seconds, to scale

All four workers launched at the same instant. Click a lane name to read that agent's actual output.

GALLERY
119.5s
BUILDS
148.6s
STAFF
160.4s
GUIDES
177.6s

Wall clock for all four: 178 seconds. Run one after another they'd have taken ~606. Parallel employees are free when the brain is local — four at once cost exactly what one costs: nothing.

the same work, single file: ~606s — 3.4× slower

Every agent's real output

Verbatim JSON each worker wrote to disk, plus the dispatcher's merge. Untouched, typos and all.

{"agent":"builds","domain":"workspace build folders","total":83,"newest":["lfm-builds","maple-tests","higgsfield","qwen38-builds","apollo-conversations.jsonl"],"note":"The workspace contains 83 build folders with lfm-builds being the most recently modified."}
{"agent":"staff","domain":"Hermes AI employees (profiles)","total":41,"names":["agents-a1","deepseek","gemma4","grok-4-5","hermes-cloud","lfm","qwen-3-7","mybot"],"note":"Selected 8 interesting profile names from the Hermes AI employee directory."}
{"agent":"guides","domain":"published guide pages","total":313,"newest":["qwen-agent-os","hermes-free-forever","maple-preview","hermes-0-20","hermes-qwen-3-8"],"note":"Total of 313 guide pages found, sorted by newest first"}
{"headline":"Four-agent swarm closed out with builds, gallery, staff, and guide coverage","numbers":{"build_folders":83,"gallery_assets":14,"ai_employees":8,"guide_pages":313}, …} // the dispatcher wrote its JSON block twice and made two mistakes — see the review below

The reviewer's pass — checked against the disk

A swarm without review is a rumor mill. Every claim was re-run against the filesystem:

builds · 83✓ exact — 83 entries on disk
staff · 41✓ exact — 41 profiles on disk
guides · 313✓ exact — 313 pages on disk
gallery · 5 + 9✓ videos exact · images off by one (9 vs 10 on disk)flagged, corrected in the headline stats above
dispatchersaid 8 employees (grabbed the sample list, not the total 41) and duplicated its JSON keysboth caught here

Three of four workers were exact on the first pass. The reviewer lane exists for the rest — same Planner → Builder → Reviewer discipline the Agent OS runs everywhere else.

What this is: a real recorded run of LiquidAI's LFM2.5 (2.6B, on-device) working as a six-run agent team through Hermes on one Mac — timings from the process clock, outputs verbatim from disk, review done against the filesystem. Page styling is the Agent OS template; every number and JSON string is the swarm's own. From Run Hermes FREE Forever.