Use cases
Most “multi-agent” demos are a relay race: one agent works, hands over, waits. These run side by side. A lead agent splits the problem and launches 16, 32 or more helpers at once; each investigates its piece with real tools, and the lead merges the findings into one answer.
All on one Mac, on inferd, with open-weight Qwen3.8 models. Nothing goes to a cloud model.
Dozens of conversations generate side by side on one GPU. In our tests, 31 of 32 agents were writing at the same moment.
See throughput as agents are added →When every agent needs the same long document (a report, a regulation, a rulebook), inferd reads it once and every agent starts from that memory. Agent 32 starts as fast as agent 1, and nobody pays to read it again.
See it measured: one agent reads, 31 start from it →Timings: Mac Studio (M5 Ultra), inferd, 5 November 2026. Models: Qwen3.8-Flash-Next (the faster) and Qwen3.8-27B. Agents run at low thinking effort, OCR readers with none. A higher effort was slower and no more accurate.
The featured three
Pick a scenario for the story, how it ran at 16 and 32 agents on both models, and how it could be better.
32 auditors, one 40,000-token report
A water utility's draft annual report (1,640 lines, about 40,000 tokens) must be checked before publication. A lead auditor splits the work by specialism (totals, dates, citations, regulatory references, units, consistency) and by section. Verifiers use a calculator and regulation and citation lookups; the lead merges the findings into one report with a chart. Twenty problems are planted.
Every verifier needs the whole report: page 40 can't be checked against page 3 without both. Without a shared cache that is 32 readings plus a re-read every turn. With inferd the report is read once and pinned. This is where the cache does the heavy lifting.
At 32 agents the swarm handled 6.5 million input tokens and 99% came from shared memory. The machine spent its time thinking, not re-reading.
Lower is better
Higher is better
Total agent working time ÷ clock time.
Higher is better
The share of the text the agents were given that came from the cache.
| 16 agents · Flash-Next | 32 agents · Flash-Next | 32 agents · Qwen3.8-27B | |
|---|---|---|---|
| Time | 5 min 16 s | 8 min 19 s | 10 min 9 s |
| Generating at once | 15 | 31 | 30 |
| Parallel factor | 8× | 15× | 13× |
| From cache | 98% of 3.7 M tokens | 99% of 6.5 M tokens | 98% |
| Planted problems found | 19 of 20 | 16 of 20 | 18 of 20 |
32 analysts, one market wrap, about two minutes
A strategist launches one analyst per instrument (indices, stocks, currencies, commodities, crypto). Each pulls live prices, checks the news if its instrument moved, and draws a chart; the strategist ranks the moves in a market wrap.
Thirty instruments are thirty independent jobs: queued in sequence, run here at once, so the wrap arrives while the news is fresh. Little shared text; the cache is not the point.
Twice the analysts took only 43 seconds longer (1 min 34 s to 2 min 17 s): close to twice the work in the same window.
Lower is better
Higher is better
Total agent working time ÷ clock time.
| 16 agents · Flash-Next | 32 agents · Flash-Next | 32 agents · Qwen3.8-27B | |
|---|---|---|---|
| Time | 1 min 34 s | 2 min 17 s | 2 min 39 s |
| Generating at once | 15 | 31 | 31 |
| Parallel factor | 9× | 19× | 20× |
| Output | 15 charts and a wrap | 31 charts and a wrap | 31 charts and a wrap |
a road trip planned by 32 scouts
“Plan a three-day drive from Lyon to Marseille with the best food, wine and viewpoints.” A planner launches one scout per candidate town, plus specialists for food, wine, history and views. Scouts check each town, ask a real routing service how long the detour adds, and search the web; the planner ranks the stops and draws the itinerary on a map.
A traveller compares twenty towns one at a time; the swarm compares them all at once, with driving times from a routing service, not the model's memory.
Lower is better
Higher is better
Total agent working time ÷ clock time.
| 16 agents · Flash-Next | 32 agents · Flash-Next | 32 agents · Qwen3.8-27B | |
|---|---|---|---|
| Time | 1 min 25 s | 1 min 54 s | 2 min 59 s |
| Generating at once | 15 | 31 | 31 |
| Parallel factor | 7× | 12× | 10× |
| Output | itinerary and map | itinerary and map | itinerary, no map |
All nine
All nine at 16 agents, Qwen3.8-Flash-Next, Mac Studio. Switch the measure to compare.
Lower is better
Higher is better
Higher is better
Scenarios without a score on the answer key, or without shared reading to measure, are left out of this view.
Higher is better
Scenarios without a score on the answer key, or without shared reading to measure, are left out of this view.
Open a row for the detail.
A draft article about inferd is checked against inferd's own documentation, pinned once. One checker per claim, all starting from the same text.
All 9 planted errors found.
The real text of DORA, the EU's Digital Operational Resilience Act (Articles 5–30, about 22,000 tokens), is pinned beside a fictional bank. Checkers examine the bank's policies, controls, systems and supplier register against the obligations; two challengers remove false alarms; a writer produces the gap register.
11 of 12 planted gaps found.
An overnight queue of 20 security alerts at a logistics company. One investigator per alert works through logs, IP and domain lookups and threat intelligence; specialists then join the dots. Six alerts are one intrusion from one attacker's infrastructure; two are harmless traps.
10 of 10 on the answer key, traps avoided.
Sixteen alerts at a fictional UK bank. One investigator per alert checks customers, transactions, shared phone numbers and devices, and sanctions lists; analysts join the cases and a writer drafts the suspicious activity report. Five alerts are one mule ring paying a shell company whose director is on a sanctions list.
10 of 10 on the answer key.
Twenty-four scanned supplier invoices with noise, stamps and a handwritten “pay urgently” note. Vision agents (thinking off, which suits OCR) read the pages, check suppliers and purchase orders and record each invoice; reconcilers look for duplicates and changed bank details; an accounts-payable lead writes the exception report.
All 8 planted problems found, including both payment-fraud attempts.
An editor lists a story's key claims. Reporters search live news in 13 national editions (US, UK, France, Germany, Spain, Italy, Poland, Brazil, Japan, China, India, Australia, the Arab world); verifiers look for primary sources; analysts trace which reports are one wire story repeated; a desk chief separates “verified” from “one source, repeated many times”.
One machine, one server
Next
Not built, not measured. Each pushes a different strength further.
300 documents in an acquisition data room: contracts, leases, employment agreements, board minutes, many scanned. Overnight, 64 agents each take some documents and a checklist (change-of-control clauses, unusual termination rights, missing signatures, liabilities). By morning: a red-flag report with every finding linked to its page.
After a major storm, thousands of claims arrive with photos. Vision agents assess each in parallel: damage, whether the policy covers it, a rough repair estimate, signs of fraud (a recycled photo, damage that doesn't match the storm's path). Urgent cases, like uninhabitable homes, jump the queue.
Today's scenarios answer a question and stop. The war room keeps going: over 20 minutes a cyber incident (or a market shock, or a supply-chain failure) delivers new evidence, such as log lines, a ransom note or a regulator's call. Each event spawns a new wave of agents; the lead keeps a living situation report and an executive can question it at any moment.
Keep going