AI Workshop: What nine AI sessions cost when you arrange them badly

For months I have been working not with one AI assistant but with several AI agents at once, on one project, each with a fixed role. It works. But not for the reasons you would expect, and it cost me more than it had to. Below are the rules that survived, and the measurements behind them. None of this is copied from a blog post. Everything was measured on my own projects, and where a number is only an estimate, it says so.

Last revised on 8 November 2026. The tools change monthly; the numbers hold for the date they were measured.

1. Usage grows with conversation length, not with work

This is the insight that orders everything else. Every working step re-reads the entire history so far. When a conversation stands at 300,000 tokens of context, one more step costs about 300,000 tokens, whether it is big or small.

The consequence is uncomfortable: merging two working steps saves more than cutting 10 kilobytes of text. Do the sum yourself: one token is roughly 1.9 bytes of text, the factor my rules use. Cutting 10 kilobytes therefore saves about 5,000 tokens per step. One step fewer saves the whole 300,000 when the history stands at 300,000 tokens.

In practice: search first, then read only the section you need. Never pour a large file, a log or a JSON dump into the conversation. Bundle independent requests into one message.

2. Every message between sessions is expensive, so there are only five kinds

Anyone running several sessions in parallel lets them talk to each other. That is where the bill explodes: every message starts a full working step at the recipient, which re-reads its whole history. That is often several hundred thousand tokens for one sentence.

My answer is a list of exactly five message types: order, done report, block, question, finding. Everything else is forbidden: acknowledgements, progress updates, „just for your information“, questions a file already answers. It goes into a file the recipient reads when needed.

The check before every send: Does the recipient have to act now? No → file, not message.

3. You calculate the point at which to compact

A conversation that grows too long has to be compacted, and that costs. Guessing this threshold is the most common mistake I have seen in my own setup. I calculate it from three measured parts:

With a ceiling of 500,000 that leaves almost twice the headroom. Setting it lower means compacting for nothing; setting it higher gains nothing. You only see that when you calculate instead of guessing.

4. More agents in parallel are not faster; at some point they deliver nothing

I set this up as a series of measurements because the rules of thumb contradicted each other. Result: with five, six and even three helper agents running at the same time, runs failed completely. They did not slow down, they failed. Since then there is a hard ceiling and a cap on reasoning effort. And a long-running agent is explicitly not considered crashed.

I still hand broad work to agents, but for a different reason than you might think: what an agent reads never lands in my own history. It saves history, not usage. Its tokens count in full.

5. A cost cap has to match the advance estimate

My research program checks before every sub-step whether the budget still holds, and it assumes 1.00 USD in advance for each search run. With a cap of 2 USD, a run therefore stopped after the first of four sub-questions. Up to then it had cost 0.09 USD.

The same job with three pre-set sub-questions and a higher cap cost 0.38 USD, took five minutes and delivered 46 confirmed out of 100 checked statements. Revising 16 articles on this site came to about 2.85 USD of research in total. A cap that is too tight saves no money; it only stops good runs.

6. Three levels of evidence, and „not found“ is not a disproof

Every researched statement carries one of three marks:

The third mark is the most important and the most misunderstood: „not found“ does not prove the opposite. A question may stay unanswered. This level recently kept me from a wrong business decision: a market gap I thought was proven had simply never been searched.

7. Checking is working time, so it gets a cap

The rule that saved me the most time is unpopular: checking time stays well below building time. Never three hours of tests for five minutes of code. Before every further test there is one question: Which decision does its result change? None → no test.

On top of that comes an order in which an AI run comes last. First a script, a comparison, a measurement. Then a local model on own hardware. Only when neither can answer, an AI run with its cost estimate stated in advance. And the other way round: whatever an AI finds that a script could have found becomes a script before the next run starts.

8. What is never capped: three checks that always happen

Saving must never cost something that does not come back. So three things are never skipped:

1. one look before anything irreversible happens;
2. one run on real data when a tool that others rely on is changed;
3. one look at a result before it is handed over as finished.

9. The rules that came from damage

These were the most expensive for me, so the damage stands next to each:

• Whatever you start, you stop. 49 forgotten browser processes in the background brought a workstation to a halt. The rule now covers the whole process tree and the temp folders, also on abort.
• Build leftovers eat the most space. A system disk stood at 97 and 99 percent twice, not because of data but because of build folders. Since then there is exactly one build folder per project, and a free-space check before every build.
• A rule file that always loads is paid for on every request of every session. So whoever adds something cuts the same amount elsewhere. Whoever finds nothing to cut writes the debt, dated, into the file. Said out loud only, it is gone at the next start.
• New text is checked against what exists. A contradiction between two rules goes unnoticed, because each rule looks right on its own. That is the nastiest error in this whole setup.

10. The rule that has nothing to do with technology

Objecting is a duty. A role must object to every instruction that is neither measured nor proven elsewhere, and it does so to the lead, not by acting on its own. Whoever decides, decides; they do not pass the decision on. And whoever acts agrees on it first. Otherwise two roles work past each other without anyone noticing, because each one is right on its own.

The same goes for results: I have my own judgements refuted and write the refutation down instead of smoothing it over. On 8 September 2026 such a counter-check turned the result around twice. That is not a misfortune, that is the purpose of the setup.

We have reviewed single AI tools separately, with prices and limits, for example ChatGPT Plus (in German) and Hugging Face (in German).

What it costs if I set it up for you

The text above is complete. There is no second version behind a form. If you want to implement it yourself, you can; it contains everything you need.

If you want it faster:

Coaching includes 4 hours a month. Unused hours carry over to the next month once, then they lapse.

This offer is exclusively for businesses; all prices net plus VAT.

What you do not get: a guaranteed saving, a state that lasts a year (the tools change monthly), or operations and on-call support.

Frequently asked questions

Can I implement this myself without a setup?

Yes. The ten rules above are complete. The setup saves you time; it does not unlock anything.

What does a first call cost?

90 minutes cost €150 net. I look at your setup and tell you what I would change.

Which AI tool do you work with?

With Claude Code by Anthropic, in several sessions with fixed roles. The rules above depend on the way of working, not on a vendor.

Do you guarantee that my AI costs will go down?

No. I show you where usage comes from in your setup and what I would change. I do not guarantee a saving.

Contact: