- Claude is a reasoning model. It's slow on repeats when you ask it to reason at every step. You need fast execution.
- On the same multi-step task, run two ways, a compiled agent finished in 1m 21s at $0.063, versus 7m 58s at $6.26 for the LLM-per-step run on Claude Opus 4.7. That's roughly 99 times cheaper and 6 times faster.
- Letting a model decide every step earns its keep the first time you run a task. By the hundredth run, most of that reasoning is wasted spend.
- The fix is to reason once, at build time, compile the repeats into code, and save Claude for judgment that is actually new each run.
- You don't have to give up Claude Code. Assign the repeat jobs to Airtop.
If you've ever asked Claude to run a simple automation, like piping social comments into a spreadsheet, you may have watched it work and wondered why it’s so slow.
Claude is a reasoning model. It's very good at reasoning. You don't need a model to reason at every step. You need fast execution. Asking it to re-plan a solved job on every run is like asking an intern to rebuild the expense-report process every time they file one.
That loop is expensive and slow because of how these loops work, not because anyone designed Claude to be slow. It re-solves the same problem on every step of every run, regardless of how strong the reasoner is. This article walks through what that loop looks like, what it costs against a compiled alternative, and what to change so Claude does the thinking you need instead of the thinking you don't.
What Claude is doing when it re-solves a job
When Claude runs a repeatable job through chat or a Computer Use style workflow, it isn't executing a script. It decides at each step: look at the current state, decide what to do next, take the action, then look again to confirm the action worked.
That reason-act-verify cycle turns a five-step task into 20 or more model calls. Each call costs tokens and time. Each call also gives the model a chance to misread the state, take the wrong action, or lose track of where it is. That produces the familiar failure modes: an empty extract, a login that silently fails, a session that dies mid-task, a result that looks plausible but is wrong.
The reasoning itself isn't bad. It runs fresh, from zero, at every step, instead of once. Repeat jobs end up here whenever there is no usable API, or whenever the agent is asked to re-plan work it already solved.
We ran the same multi-step task, run two ways. The first run used an LLM-per-step loop on Claude Opus 4.7: screenshot, reason, act, verify, repeated at every step. The second run used a compiled agent that had already reasoned through the task once and turned that reasoning into deterministic steps.
The LLM-per-step run took 7 minutes 58 seconds and cost $6.26. The compiled agent finished the identical task in 1 minute 21 seconds for $0.063, about 99 times cheaper and roughly 6 times faster, for the same result.
At one run, that gap is a curiosity. At 100 runs a day, it decides whether a workflow scales or quietly drains budget while operators wait. A GTM engineer running lead enrichment or portal monitoring on a schedule pays for the same reasoning to happen thousands of times, most of which the task didn't need.
You don't need a model to reason at every step
Agency, meaning letting a model decide every step on its own, earns its keep the first time a task runs. Nobody has solved the flow yet. A model reasoning through that ambiguity live does real work. By the hundredth run, that same reasoning is wasted spend.
Airtop's research on computer-use behavior found run-to-run variability of 20 to 60 percent on identical tasks, because the model keeps re-deciding steps that already have a known-correct answer. Agency is worth the most the first time you do something and worth the least by the hundredth. Repetition doesn't call for fresh judgment. It calls for reliable execution.
Airtop agents split that work. Repeatable sequences compile into code. Judgment that changes every run stays a reasoning call. Navigation, clicking, typing, scrolling, and file uploads are mechanical once solved. Deciding whether a lead is real, whether a page changed, or whether a human should review a record still needs a model. Code-first approaches also tend to need fewer reasoning steps than JSON-based tool calling, since the model isn't re-deriving control flow it already worked out.
Keep Claude for judgment, compile the execution
The practical shift is to reason once, at build time, then run the compiled result like software after that. Airtop calls this "deterministic AI": the model handles ambiguity exactly once, and the resolved steps get compiled into reusable code that runs the same way every time.
The build process uses a teaching loop: you chat to describe the task, watch the agent attempt it, correct the parts that go wrong, and the correction compiles into the agent's logic. Once that loop finishes, the agent stops guessing and starts executing.
You don't have to give up Claude Code. Assign the repeat jobs to Airtop. Claude can still plan and sequence. A compiled agent runs the work that is the same every night, instead of asking Claude to re-derive the same steps.
What to do about it
Start by identifying which Claude jobs run more than a handful of times. If a workflow runs once, agency makes sense. If it runs daily or on a schedule, compile it.
Airtop compiles the repeats and keeps a reasoning call where the answer is new each run. You don't have to sit there and assign each step to Claude or to Airtop by hand.
Then build the compiled version. With Airtop, you describe the workflow and Airtop compiles it into a deterministic agent, ready to run at scale, while Claude keeps doing what it's good at: planning, judgment calls, and handling cases that are genuinely ambiguous.
FAQs
Why does Claude take so long to complete repeatable jobs?
Most of the delay comes from the LLM-per-step loop, not Claude's reasoning. Claude reasons, acts, then verifies the result, and repeats that cycle for every step. A five-step task can turn into 20 or more model calls, and each one adds time and cost.
Is Claude itself the bottleneck, or is it the way the task is run?
It's the way the task is run. The same model produces very different results depending on whether it's re-deciding every step from scratch or executing steps that were already worked out once. Compare the same multi-step task, run two ways: 7m 58s for the LLM-per-step run versus 1m 21s for the compiled version.
What's the difference between an LLM-per-step agent and a compiled agent?
An LLM-per-step agent, sometimes called a computer-use agent, reasons about every action live, every single run. A compiled agent reasons through a task once, then runs the resolved steps like software after that, calling the model only where the work still needs judgment.
Does compiling an agent mean giving up Claude's reasoning?
No. Compiling moves the model's judgment to where it's needed: reading a page and deciding whether a condition is met, or extracting data across pages that change shape. Navigation, clicking, and typing don't need fresh reasoning once they're solved, so those steps compile.
How much faster and cheaper is a compiled agent than a Claude-driven loop?
On the published benchmark, a compiled agent finished in 1m 21s for $0.063, versus 7m 58s for $6.26 running the same task with Claude Opus 4.7 in an LLM-per-step loop. That's roughly 99 times cheaper and about 6 times faster, for the identical result.
Can I still use Claude Code and get compiled-agent speed?
Yes. Keep Claude Code. Assign the repeat jobs to Airtop. Claude can still plan and sequence. A compiled agent runs the work that is the same every time.
Why does my automation end up in a browser loop instead of using an API?
Many vendor portals and legacy tools have no usable API, so the only way in is a real browser session, clicking through the interface the way a person would. That's often the actual reason a workflow ends up leaning on Claude to drive a browser step by step in the first place.
Does this mean agency is never worth it?
No. Agency, letting a model decide freely, is worth the most the first time a task runs, when nobody has solved the login flow or field mapping yet. It's worth the least by the hundredth run, when the task is the same and the reasoning just repeats.
Try Airtop For Free
If your Claude jobs run more than once, the fix isn't a faster model. It's compiling the repeatable steps so Claude only handles the judgment that still needs it. Spin up your first agent in five minutes and try it for free.
