Claude Code Parallel Agents: 20 Sessions From a Conference

I ran 20 Claude Code parallel agents from a conference seat. Here is the record, how I decide what to parallelize, and why review is the real limit.
I showed up at a conference where the air conditioning was cranked to maximum, wearing shorts by mistake. So I ran Claude Code at 20 parallel sessions — video editing and similar work — and joked that the heat was my heater. The heater part was a joke. What I actually did, and what I want to record here, is that 20 parallel agent sessions ran as normal work from a conference seat. Below is the record plus the criteria I use before I parallelize anything.
What happened: 20 Claude Code parallel agents from a conference seat
The facts are simple.
- The venue’s air conditioning was at maximum
- I had dressed wrong and come in shorts
- On the spot, I ran Claude Code at 20 parallel sessions doing video editing and other work
- I posted that I was using AI and GPUs as heating while still getting work done
- Someone in the thread replied that I would not need a heater; I said it was a trick
- I added that I should have just worn long pants
I want to be clear about the joke. Yes, hardware gets warm under load — everyone knows that. I did not measure a temperature, I did not test it as a heating substitute, and I am not recommending it as one.
But strip the joke out and something stays: work was flowing across 20 parallel sessions while I sat in an audience seat. A few years ago that was not a thing I could do. The heat got the attention; the change underneath it is the part worth writing down.
One note on the Japanese context, since this post traveled mostly in Japanese. Over-air-conditioned event halls are a running joke here, and so is the self-deprecating apology at the end of a boast. That mix — a flex wrapped in an apology — is a big part of why the post landed. If you are an overseas company trying to read Japanese AI discourse, that tone matters as much as the content.
The real limit on parallel agents is how you cut the work
The ceiling on parallel sessions is not set by your machine. It is set by how far you can cut the work into independent units. I only fan something out when three things are true at once.
- The units are independent. Each one can start without waiting on another’s result.
- I can judge completion myself. Looking at the output tells me whether it is good.
- Failure does not propagate. If one session goes wrong, the others are not wasted.
Video editing fits all three. It splits by clip, by cut, by asset, and a quick look at the output settles whether it is usable. What does not fit is a single long chain of dependent judgment — where each conclusion becomes the premise of the next step. Force that into parallel lanes and you will redo all of it later.
So before increasing the count, the question is not “can my setup handle 20?” It is “does this work actually divide into 20?” If it does not, parallelizing only adds waiting and rework. The 20 was not a target. It was just how many independent pieces that day’s work happened to have.
Review is the bottleneck, not output
Run 20 sessions and 20 results come back. At that point, what matters is not production speed but intake speed. My practical ceiling on parallelism is how fast I can look at output and decide what to accept.
Push past that and the failure mode is predictable.
- Output piles up that nobody has read
- Good and bad results advance together, mixed
- You lose the thread of where the approach changed
So before raising the count, I do three things.
- Write down the acceptance criteria before dispatching. If I cannot say what “good” looks like in a sentence, the session is not ready to run.
- Standardize the shape of the output so results are comparable side by side instead of each needing its own reading.
- Block time for review as one batch rather than interrupting myself twenty times.
That sequence alone changes how much actually ships at the same session count. Which leads to the uncomfortable conclusion: people who are good at parallel agents are not the people who write clever prompts. They are the people who can decompose work and articulate a standard for “done.” That stays a human job.
A conference seat becomes a workplace
Time spent traveling to events and sitting in sessions used to be reading-and-replying time for me. Work that consists of dispatching instructions and waiting moves forward from a seat, which quietly changes what counts as downtime. Freezing in shorts while 20 sessions ran turned out to be a test of exactly that.
If you do this in the field, a few things need deliberate handling. Assume you are responsible for your own network and power — venue conditions are not a plan. Vary what kind of material you are willing to work with by location. Whether a file is appropriate to open with people around you is a separate judgment from how many sessions you are running. Getting careless there turns the convenience into a liability fast.
What this means for Japanese companies — and for overseas AI companies
If you bring this way of working into a Japanese company, do not lead with the number. The first question here is rarely “how fast?” It is “who reviews the output, and who is accountable for it?” Twenty parallel sessions is an eye-catching figure, but as the centerpiece of an internal pitch it reads as “this person is not in control.”
The order that actually gets approved starts from the review mechanism. What standard decides accept or reject. Who looks at it. Where the record lives. Once that is settled, the session count can grow as much as you like.
I think the same holds for overseas AI companies communicating parallel-agent capability to a Japanese audience. Positioning built purely on speed and scale lands poorly here. Reviewability, the ability to stop something mid-flight, and a trail that remains afterward — put those forward and enterprise evaluation moves. What is being assessed is less the headline number and more whether the manager on the ground can explain the setup to their own organization.
Putting it to work
You do not need to aim at 20. The order I would suggest:
- Start at two or three parallel and feel the review load yourself
- Inventory which tasks genuinely split into independent units
- Write your acceptance criteria down as a short paragraph
- Raise the count only once you can see review is keeping up
It was a joke post, but one lesson survives it. How much you can run in parallel with AI is set by how well you can divide the work. Next time I am wearing long pants.
FAQ
Does running Claude Code at 20 parallel sessions really work as a heater?
That part was a joke. Devices get warm under load, but I never measured anything and I did not test it as a substitute for heating. The only thing I actually confirmed is that 20 parallel sessions ran fine from a conference seat.
What were the 20 parallel sessions doing?
Video editing and similar work. I did not post a full breakdown, so I cannot add detail beyond that. They were tasks that split cleanly into independent units.
Is a higher number of parallel agents always better?
In my experience, no. The practical ceiling is how fast you can review the output and decide what to accept. Raise the count past that and you just accumulate work nobody has looked at.
Based on a post on XOriginal post ↗