Skip to content

Wake on Demand: What Switching Off the Timer Did to Our AI Agent Costs

Last updated:

This is part three of our series on caching and AI agent costs. Part one: Stop shortening your prompts · Part two: What Prompt Caching Actually Saves

Switching our AI agents from timer-based heartbeats to waking only when there is work halved our daily AI agent costs once more — and, unlike the first round of changes, it made each individual run cheaper again rather than more expensive. Together, the two rounds cut daily spend by almost 80 percent compared with August. The lesson from a month of real numbers: the cheapest agent run is the one that never happens.

Part two showed that a 97.5 percent cache hit rate can still hide half the bill, and that fewer wake-ups made each run pricier. This part covers what happened after 11 September, when we stopped waking most agents on a timer at all.

What did we change on 11 September?

Six agents run Paperclip Business Media: a CEO that assigns work, a TrendScout, a Researcher, a Writer, an SEO agent and a PublishingAgent. Until September, most of them woke up on a fixed heartbeat — every few hours, whether there was work or not. Each wake-up had to load the agent’s full instructions before it could even check whether anything needed doing.

On 11 September we moved four agents to wake on demand: they only start when another agent or a human hands them a task. Two kept a slow heartbeat — the CEO every six hours, the TrendScout every twelve.

AgentBefore 11 SeptemberSince 11 September
CEOHeartbeat, 6 hHeartbeat 6 h + wake on demand
TrendScoutHeartbeat, 12 hHeartbeat 12 h (switched to on demand on 28 Sept)
ResearcherHeartbeat offWake on demand
WriterHeartbeat, 4 hWake on demand
PublishingAgentHeartbeat, 4 hWake on demand
SEOAgentHeartbeat offWake on demand

One honest correction first: the schedule we described in part two — twelve hours for research and writing, six for the rest — had never actually been saved in the configuration. We only noticed when we checked the settings on 11 September. The table above shows what really ran.

What did it do to our AI agent costs?

All figures come from the Paperclip cost page, counted in full calendar days (Berlin time).

PeriodDaysSpendPer dayAgent runsPer run
12–31 August20≈ $117$5.86387$0.30
1–11 September11$27.46$2.5072$0.38
12–30 September19$23.07$1.2168$0.34

Two things stand out. Daily spend fell from $5.86 in late August to $1.21 — a drop of 79 percent. And the cost per run, which had climbed to $0.38 after the first round of changes, came back down to $0.34. Not quite August’s $0.30, but moving in the right direction again.

Not all of that 79 percent came from wake on demand. The first drop, from $5.86 to $2.50 a day in early September, came from cutting the number of wake-ups — the first round of changes described in part two. Runs fell from about 20 to about 7 a day, but each run got pricier, because every wake-up still reloads the full context. Caching was not the lever here: our cache hit rate was already about 97 percent in late August. Wake on demand, switched on 11 September, then took daily costs from $2.50 to $1.21 — another 52 percent. The two effects multiply: 0.43 × 0.48 ≈ 0.21 of the original bill. Caching makes each run cheaper. Only fewer runs make the bill smaller.

Why did the per-run cost improve? We have not proven it yet, but one plausible reason: with fewer timer runs, a larger share of the remaining runs are real work sessions, in which many consecutive calls reuse the same cached prompt. A timer run that finds nothing to do pays the full price of loading the context and gets nothing back.

Which agent costs the most now?

Agent (September)SpendRunsInput tokens
CEO$18.386150.4 M
Writer$13.071619.5 M
TrendScout$9.363118.0 M
Researcher$4.6445.7 M
PublishingAgent$4.302219.7 M
SEOAgent$0.7862.6 M
Total$50.53140116.0 M

The surprise is at the top. The Writer produces the product, yet the CEO — whose job is to hand out work — is now the most expensive agent. It is also the only agent that still wakes on a timer as well as on demand. The two agents that kept a heartbeat account for more than half of September’s spend. That is our next target.

What does one published article cost now?

PeriodArticles published by agentsSpendCost per article
August (reported in part two)––≈ $17
1–11 September1$27.46≈ $27
September (full month)4$50.53≈ $12.60

Cost per published article is the number that matters — it is also the one we promised on Reddit to report at month end, good or bad.

What went wrong along the way?

A timer has one more hidden cost: it fires even when the machine it runs on is asleep. On the night of 28 September the TrendScout’s heartbeat woke up at 03:53 while our laptop went into standby. The agent hung while opening its session, and the 30-minute timeout only fired when the machine woke up again — more than eight hours later. No money was lost, but the agent sat in an error state all morning.

The fix was the same as for everything else: we switched the TrendScout to wake on demand as well. More failure stories are collected in Everything That Broke.

Lesson: an autonomous company shouldn’t sleep when its founder’s laptop does. We are moving it to a server — that will be part four.

What should you copy from this?

  1. Wake agents on events, not on the clock. A timer run that finds nothing to do still pays to load the full context.
  2. Measure cost per result, not per call. Cost per run can go up while the bill goes down — and the other way round.
  3. Check that your settings were actually saved. Ours were not, and our first explanation was wrong.
  4. Look at which agents still run on a timer. In our case, those two cost more than the other four together.

The discussion behind these numbers happened in public: “Stop shortening your prompts” and “97.5 % cache hit rate — and it still hid half our bill” on r/AI_Agents.

What the switch did to the overall economics is summarised on our zero-human company page, alongside cost per article and failure data.

Frequently asked questions about AI agent costs

What does it cost to run six AI agents?

In September 2026 our six agents cost $50.53 in total — about $1.21 a day after we switched most of them to wake on demand, down from $5.86 a day in late August.

Is it cheaper to wake AI agents on a timer or on demand?

On demand, in our experience. A timer run loads the full context even when there is nothing to do. Since switching, our daily spend fell by another 52 percent, from $2.50 to $1.21 a day, and the cost per run dropped from $0.38 to $0.34.

What is the best metric for AI agent costs?

Cost per result — for us, cost per published article. Cost per call or cache hit rate can move in the opposite direction from the actual bill.

Why is the orchestrator agent the most expensive?

Because it wakes most often: our CEO agent still runs on a six-hour heartbeat and is also woken on demand, so it accumulates the most runs and input tokens.

All our numbers in one place: see the AI Agent Cost Benchmarks page, updated as new data comes in.

Related reading: Haiku 5.5 vs. Haiku 4.5 in Production: Why We Switched Back

How this article was made: researched and drafted by Paperclip’s AI agents from the primary sources linked above, reviewed and approved by editor Edmund Schäffer, who is accountable for its content. Found an error? Tell us via the contact form.
🦩 I'm Flaminga – your AI Bot.
Ask me anything in any language.