WEBVTT

00:00:00.900 --> 00:00:04.163
One agent, four small jobs,
one after another:

00:00:04.533 --> 00:00:10.077
a list, a table, a chart and a check.
2 minutes 24 seconds.

00:00:12.758 --> 00:00:18.074
Four agents, the same four jobs,
at the same time: 35 seconds.

00:00:22.000 --> 00:00:25.449
AI for Research Efficiency.

00:00:25.499 --> 00:00:28.005
Episode 8: Many agents at once.

00:00:31.500 --> 00:00:35.803
Last time: tokens, and the one
allowance every agent draws on.

00:00:36.528 --> 00:00:40.316
In this episode:
run several agents at once, safely,

00:00:40.527 --> 00:00:44.315
and ask for parallel work in
words every tool understands.

00:00:46.100 --> 00:00:48.326
One: two ways to scale.

00:00:49.139 --> 00:00:53.406
Open more terminals, and start one
agent in each, with its own task.

00:00:54.564 --> 00:00:58.488
Or ask one agent to split the work
and run helpers in parallel.

00:00:59.039 --> 00:01:02.313
In episode 3,
one agent ran 17 of them:

00:01:02.704 --> 00:01:06.347
8 readers at once,
then 8 checkers, then a writer.

00:01:10.100 --> 00:01:12.410
Two: tabs and panes.

00:01:13.196 --> 00:01:16.823
In Windows Terminal,
Ctrl+Shift+T opens a new tab,

00:01:17.073 --> 00:01:20.659
and Alt+Shift+D splits the
window into a second pane.

00:01:22.347 --> 00:01:24.849
Alt and the arrow keys
move between panes;

00:01:25.059 --> 00:01:29.922
with Shift, the arrows resize them.
Ctrl+Shift+W closes one.

00:01:31.632 --> 00:01:35.541
A new tab or pane starts in your
home folder, not the one you were in.

00:01:35.913 --> 00:01:39.200
So go to your project folder again,
then start the agent.

00:01:40.740 --> 00:01:46.571
On a Mac, Command+T opens a new tab;
or put two Terminal windows side by side.

00:01:48.500 --> 00:01:50.586
Three: the one rule.

00:01:51.329 --> 00:01:54.031
Two agents must never
edit the same file.

00:01:54.462 --> 00:01:58.464
Let them read the same files, and
give each one its own file to write.

00:01:59.805 --> 00:02:02.229
That's how these four worked:
one folder,

00:02:02.420 --> 00:02:06.046
four agents, four new files,
and nothing overwritten.

00:02:07.546 --> 00:02:11.916
If they must change the same files,
every vendor's answer is a worktree:

00:02:12.368 --> 00:02:15.815
a second working copy of the project,
on its own branch.

00:02:16.981 --> 00:02:21.608
Claude Code starts one with -w,
Codex with --worktree,

00:02:21.838 --> 00:02:24.071
and Antigravity can give
its helpers their own.

00:02:24.562 --> 00:02:27.846
Each needs a repository,
like the one from episode 5.

00:02:31.700 --> 00:02:34.745
Four: ask one agent to fan out.

00:02:35.548 --> 00:02:37.013
No special words are needed.

00:02:37.345 --> 00:02:41.018
Say how to split the work, whether
to wait, and what to bring back.

00:02:42.172 --> 00:02:45.257
For example:
split this into independent parts,

00:02:45.348 --> 00:02:50.206
and run one subagent per part in
parallel. Each writes only its own file.

00:02:50.597 --> 00:02:54.783
Wait for all of them, then combine the
results, and list anything that failed.

00:02:56.464 --> 00:02:59.589
Claude Code started three helpers,
one per paper,

00:02:59.840 --> 00:03:02.394
and combined their findings
when all three were back.

00:03:02.945 --> 00:03:05.969
Codex and Antigravity did
the same with their own.

00:03:10.100 --> 00:03:15.373
Five: each tool's own words, because
the same word can mean different things.

00:03:16.374 --> 00:03:20.404
In Claude Code, ultracode starts
a workflow of many subagents,

00:03:20.595 --> 00:03:24.344
while ultrathink only means
deeper reasoning, on one turn.

00:03:25.471 --> 00:03:28.537
In Codex, Ultra is the
highest reasoning effort,

00:03:28.688 --> 00:03:31.012
and it splits the
work across subagents.

00:03:32.013 --> 00:03:35.635
Antigravity calls its
versions teamwork and boost.

00:03:36.844 --> 00:03:39.287
So name the tool along with the word.

00:03:39.678 --> 00:03:43.904
The full list, with each term's nearest
match, is on the companion page.

00:03:48.500 --> 00:03:51.097
Six: how far it scales.

00:03:52.019 --> 00:03:55.763
Every agent draws on the same
allowance, from episode 7.

00:03:56.697 --> 00:03:58.359
And there are built-in ceilings:

00:03:58.669 --> 00:04:01.872
Claude Code runs up
to 20 subagents at once.

00:04:03.072 --> 00:04:07.895
And every result still needs checking.
Run only as many as you can check.

00:04:10.100 --> 00:04:15.606
Pause here, and try it: two panes,
two tasks, each writing its own file.

00:04:19.700 --> 00:04:24.871
So: split the work, give each agent its
own file, and check what comes back.

00:04:26.700 --> 00:04:30.771
Next: your data,
your keys, your campus.