NAME
ETOOMANYTHINGS?Run fewer agents
Josh Bleecher Snyder co-founded exe.dev, which gives developers and their coding agents their own Linux machines. At a conference, he sat through a bunch of software factory demos, and they were all about task management. "Why is everyone building task management and calling it a software factory?"
- Snyder calls task management a band-aid for running too many agents.
- His call to arms is to "make it possible to be equally productive with fewer agents."
- The title reads like a Unix error code, and the closest real one is
EMFILE, too many open files.
CAUSE
A slow agent is an invitation to start another
Snyder traces the task boards to latency. Agents are slow, and "the obvious, easy fix to latency is to hide it by starting a new agent every time you get blocked."
Faster models will shrink the wait eventually. Snyder doesn't want to wait for them.
- More agents shorten the wait, but one person still reviews every result in order.
- Snyder's line is that concurrency "trashes flow state and thrashes our mental page caches."
- Switch Lab A from 3 slots to 6, and agent work ends at 8 while the last review stays at 26.
| SLOTS | AGENTS DONE | ALL REVIEWED | REVIEW |
|---|---|---|---|
| 3 | 16 units | 26 units | 18 units |
Going from 3 slots to 6 finishes agent work at 8 units instead of 16. The last review still ends at 26.
WORKAROUND 1
Reportedtried at exe.dev
Let a fast model do the talking
A powerful model starts the task, and Luna 6 handles the conversation, with "no coding tools" and a context window kept "intentionally short and focused on what the human conversation requires."
"This works reasonably well," Snyder writes, but chat still squeezes everything through a tiny pipe. "Luna might be an exciting, bendy straw, but it's still a straw."
- Snyder's first fix splits the talking from the working, with a fast model in front of the slow one.
- Luna is back from the harness story, where a harness made the budget model dependable.
- Watch the straw in Lab B. Do quicker replies make up for a narrow pipe?
WORKAROUND 2
Reportedshipped in Shelley
Talk and point instead of typing
Most coding harnesses show the person "a single linearized text thread, typically interspersed with lots of irrelevant tool call noise." Snyder wants to "optimize for human I/O" instead.
"I can still pick a Go panic stacktrace out of terminal logs scrolling by at 30fps," Snyder writes. Show-and-tell ships in Shelley, exe.dev's coding agent, and Snyder finds the feedback richer than chat.
- Eyes read structure fast and talking beats typing, so Snyder's second fix records both.
- Shelley transcribes the audio with word-level timestamps and matches each word to what was on screen.
- Press Play in Lab C and follow the highlighted word as the frame changes.
Keep the approval step
spoken
Keep this approval step before anything goes live.
"this approval step" points at the publishing rule.
WORKAROUND 3
Reportednot shipped yet
Ask the agent for a page instead of a thread
Snyder's prototype page holds decisions, open questions, examples, and git diffs. A record button captures his voice while a DOM tracker logs what he views, points at, highlights, and clicks, with timestamps.
"I find that this reduces mental overhead; I am freed to focus on the content."
- For work that isn't visual, Snyder has the agent answer with a page instead of a thread.
- His line is "Agents are happy to be told, but humans prefer to be shown."
- Click through Lab D. An agent helped build this page too, minus the record button.
tail -f agent.log
review.html
requireApproval();
publish();Both layouts hold the same decision, but only the page gives it a fixed place.
EXPERIMENTAL
Reportedstill an experiment
Let the agent answer while you're still reading
Reading a page well takes time, so Snyder puts that time to work. The agent's replies stream into a tab on the page he's reading.
"By the time I've spent 10 minutes reading and probing, I will inevitably have asked questions and provided provisional guidance."
- Snyder's newest experiment sends prefixes of his feedback to the agent while he's still recording.
- Some answers arrive before he switches away, which he says "sneaks in an extra round trip."
- Step through Lab E. How does the agent tell a half-finished thought from a final one?
You read the proposed change
You ask whether the approval check covers every way to publish. You're still talking.
A half-finished thought can change, so the agent needs to know which feedback is final. The caveat is ours.
BUGS
He's getting his attention back, a little at a time
"I am not yet at a point where I have the same net throughput with one or two agents that I do with many, but my attention is being restored to me, little by little."
- Snyder hasn't matched his many-agent throughput with one or two agents yet.
- In the harness story, Naresh Jain's cost figures left out the person reviewing the agents' work, and Snyder's experiments start from that person.
- Would you give up some throughput to get your attention back?
| idea | status |
|---|---|
| fast model for conversation | tried at exe.dev |
| show-and-tell in Shelley | shipped |
| HTML page with DOM recording | not shipped yet |
| streaming partial feedback | still an experiment |
| same throughput, fewer agents | not reached yet |
SEE ALSO
- etoomanythings(exe.dev) Josh Bleecher Snyder, October 1, 2026. Checked October 19, 2026.
- show-and-tell-shelley(exe.dev) Shelley's show-and-tell announcement, linked from the essay.
The labs are illustrations, and nothing on this page records. Status follows the October 1 essay.