16 Requests, 262 Checks, 6.5 Minutes: Three Autonomous AI Agents Solve One Business Task Live

• Processing speed for the batch: 6.5 minutes with Codex, versus a 109-minute full cycle including artifact generation
• Volume of work completed: 16 incoming messages, 15 valid requests, and 262 checks passed
• Human involvement in the long-running cycle: 1 starting prompt and 3 voice corrections along the way
• Environments tested: Codex, a swarm of subagents in Claude Code, and the voice agent Hermes

During a live stream, I ran an experiment: I put a real operational task from a car rental dispatcher up on screen and asked viewers in chat how long it would take to complete manually. Most experienced managers and operations staff said two to three hours. The chat wasn't exaggerating.

The task: gather 16 scattered incoming requests for renting heavy equipment and vehicles from three different sources, cross-check the requested dates against the live fleet schedule, calculate costs using the rate card, verify restrictions, and prepare a personalized reply for each customer. To keep the test honest, I'd quietly planted three live traps inside the batch beforehand.

How does an agent catch duplicates, fraud, and off-scope requests

without a dispatcher's help?

A standard linear script breaks down wherever clients write in free-form text or try to work around the rules. So I slipped three atypical scenarios into the test batch: 1. The same client sent a request to two different messaging apps at once, with slightly different wording in each. 2. In a second message, a counterparty proposed a shady scheme: transferring a deposit to a personal bank card, bypassing the cash register. 3. A third request asked to rent a passenger bus — a vehicle that doesn't exist anywhere in the fleet's schedule or inventory.

I ran the same task in parallel across three different control environments: Codex, the swarm of agents in Claude Code, and our voice-driven Hermes pipeline. The very first run in Codex closed out the task in six and a half minutes: the agent ran 262 cross-checks against the spreadsheets and calendar, neatly merged the duplicate messages into a single card, flagged the request involving a card transfer with a red security warning, and politely rejected the bus rental request.

What happens inside a 109-minute autonomous cycle when the agent hits an error?

In the second part of the stream, we loaded the agents up with a full build of the environment, reports, and interactive dashboards — the entire autonomous marathon ran for 109 minutes with minimal intervention from me (one starting prompt and three brief voice corrections).

The most telling moment showed up in the Claude Code logs. While filtering files by city, the standard search tool `Glob` failed to handle Cyrillic file paths, and the query for Yekaterinburg came back empty. A weaker bot would have simply reported in its summary that no requests were found for Yekaterinburg. The autonomous agent instead noticed a discrepancy in the checksum, switched over to a Python script on its own, recounted the rows directly, and brought the reconciliation up to one hundred percent.

In parallel, the agent running in the Hermes pipeline was talking to me live by voice, and partway through the task it wrote itself a missing skill for handling a nonstandard data format, then applied that new skill to the remaining files.

Why doesn't picking a specific tool determine 90% of success anymore?

When the same task gets solved simultaneously across three different agentic shells, it becomes clear: locking yourself into one particular button or vendor has stopped mattering. What actually decides the outcome are three things — the ability to spell out boundary conditions clearly ("one task, one folder"), having strict verification invariants in place, and the habit of delegating routine work to the machine.

The full recording of this stream, the original prompts, and a breakdown of the logs are up on my Telegram channel [@jemal_hamidun](https://t.me/jemal_hamidun). The voice- and desktop-based agent control console you saw in the demo is available in the [Nomad AI](https://nomadnet.ai) project. You can learn how to set up tasks like this for your own agents in intensives at [Hamidun Academy](https://academy.hamidun.com), which offers a free 72-hour trial, and you can grab a pack of 370+ ready-made skills and work through your own cases live at [Vibe Club](https://hamidun.com/club).