Last week I needed current financials for 50 companies. Revenue, EBITDA margin, net debt. Into a spreadsheet. Sourced.

I did what most people do: I asked Claude to go do it.

It started strong. Companies 1 through 12 were clean. By company 30 it was getting sloppy — a margin that didn’t tie, a “2024” figure that was actually a 2023 restatement. By company 45 it had quietly stopped citing sources altogether.

It didn’t get lazy. It got full.

Here’s the part nobody explains. An AI has a fixed amount of working memory — a context window. Every filing it reads, every page it fetches, every intermediate note it scribbles, all of it stacks up in that window. A 50-company research task doesn’t just take a long time. It slowly buries your original instructions under 400 pages of 10-K.

Subagents fix that.

What a subagent actually is

A subagent is a fresh instance of Claude with its own clean context window, its own instructions, and exactly one job. It goes off, does the work, and hands back only the answer. The 50 filings it read to get there never touch your main conversation.

It’s the difference between reading 500 pages yourself and asking an analyst to read them and give you the summary.

The 50-company job, redone

Here’s what I actually ran in Cowork:

1. Split the work. “Break these 50 companies into 5 batches of 10. Spin up a subagent for each batch.”
2. Brief each one precisely. Every subagent got the same instruction: Find the most recent reported revenue, EBITDA margin, and net debt for these 10 companies. Primary sources only — 10-K, 10-Q, or IR page. Return a table with a source URL for every figure. If you can’t verify a number, write UNVERIFIED. Do not estimate.
3. Run them in parallel. All five worked at once. What had been a 90-minute grind finished in about 15.
4. Add a checker. One more subagent, fresh eyes, one job: independently re-pull a random 10 of the 50 and flag anything that doesn’t match. It caught two errors.
5. Assemble. The main thread took five clean tables and one QA report and built the spreadsheet.

My main conversation never saw a single filing. That’s the whole trick.

Where this breaks

Subagents can’t talk to each other. So don’t use them when the work needs shared judgment: five agents writing five sections of one memo will hand you five different voices and three contradictions.

And a vague brief gets you vague results, times five. The level of detail you’d give a smart new hire on their first morning is roughly right.

The reframe

I’d been treating AI like a tool. Subagents made me treat it like a team. The skill that matters stopped being prompting and started being delegating. Scope the work. Write the brief. Check the output.

Which, if you’ve ever run a team, you already know how to do.

What’s the task you’ve been avoiding because it’s too big for one pass?

Alex