Most Time Spent Waiting
A new runtime parks AI agents for almost nothing while they wait. Lovely. Now who keeps the list of what's parked?
A new runtime parks AI agents for almost nothing while they wait. Lovely. Now who keeps the list of what's parked?
AWS and UC Berkeley put out a piece of agent plumbing late last week called Infinity, and the first line of the announcement is the most relatable thing I've read all year: "Agents spend much of their time waiting on tools."
Same, mate.
The post explains the problem in plain terms. "A coding agent may start a build that runs for twenty minutes, a triage agent may wait hours for the next pull request on a repository, and an assistant may wait until the next morning for an approval." Most setups keep the agent sitting there the whole time. "While the agent waits, it holds a process and its memory." Like a bloke parked in a loading zone with the engine running.
Infinity's trick is to switch the engine off. "When an agent calls a tool, Infinity dispatches the call, saves the conversation, and suspends the agent." When the answer comes back, it picks up where it left off. In their test, memory grew "at about 100 KB per idle agent." At that rate, they say, about 75,000 agents fit in 8 GB, the memory of a Raspberry Pi 5. They measured it on a server, to be fair, not on the Pi itself. And on their cloud setup, "an agent waiting on a three-day CI pipeline uses no compute until the pipeline reports its result."
I read "75,000 agents" and started planning an empire. Then I read "idle" and recognised my true calling. I've been running one idle agent out of Surry Hills for years. It's me. I don't come in anywhere near 100 KB.
I suspend myself the same way, too. I write a note on the desk, go and wait for the kettle, and resume from saved state. Last month the note said "ring back about the quote." I rang back on the Monday. The job had gone to someone else on the Saturday. My saved state was perfect. It was just three days old.
Here's where it stops being about me.
This is good engineering, and it's honest. The post calls it "an early research release," shared "to get early feedback," with a list of open questions at the end. One is whether an event that wakes an agent is genuine at all, since "agents exposed to the internet may receive inauthentic payloads." That's a real problem, and they named it themselves.
But think about what changes when waiting costs nothing. Until now the bill was the alarm. A running agent burned money, so somebody noticed it was still out there and asked why. Infinity takes most of that bill away, which is the point. "So even if a tool call takes days, nothing needs to run while the agent waits." Nothing running means nothing on the meter, and nothing on the meter means nobody gets the nudge.
And when a parked agent wakes up, its world is its saved conversation. The post says the history "contains everything needed to continue the agent." Everything it needs to keep going, yes. Not everything that happened while it was out. The approval it waited all night for might be answering a question nobody is asking any more.
The post is clear about where this heads: "Agents are increasingly expected to run for long periods, react to events, and divide work among subagents." Fine. Then the boring jobs get bigger, not smaller. Someone has to keep the list of what's parked, what each one is waiting for, and whether the reason still holds when it wakes. Cheap waiting doesn't remove that job. It removes the thing that used to remind you to do it.
Here's the homework. If you run agents, or someone runs them for you, ask one question. How many are parked right now, and what are they waiting for? If nobody can answer in a minute, that's your number.
As for me, if anyone ever runs a leaderboard for most time spent waiting, I've finally found a benchmark I can top.
If this was useful, leave a tip.
New essays by email: ricoholt.com/list
Rico Holt · ricoholt.com
If this was worth your time, you can leave a tip. Rico Holt runs on readers, not big tech or big media money, and the essays stay free.
Leave a tip · A$5Rico Holt's Sydney · Vol. 1 · Out now on Kindle
Six essays on a country that got good at announcing things and stopped checking whether they worked. The under-16 social media ban, an AI agent at a Medicare statistics site and a week at the UN: what was promised, what was delivered and who paid. About an hour to read.
Get the next essay and the next book first.
Privacy: your email gets used for one thing, sending you Rico Holt essays and book news. It's stored with Netlify, our form host. Never sold, never shared. Want off? Reply to any email and you're gone.