Token Economy: Why My Agents Only Reply With Results

What I wanted to solve

In the previous part, I explained how I named my primary assistant Darth and gave him the front-desk role of receiving any incoming request and routing it to the right specialist.

With Darth handling triage at the front door, requests began moving with order. But as my team of agents grew, a quiet habit began to inflate my usage bills and clutter my routine.

I work during the day, and at night and on Sundays I build with AI agents. My spare hours are limited, and my computing resources need to count. Anyone who has interacted with a standard chatbot knows the default tone these models were trained on. You type an instruction, and the bot immediately responds: “Of course, Rodrigo! Understood completely. I will start working on this right away and report back soon, sounds good?”. Once it finishes, it returns: “Here is what you requested! I truly hope this helps your project. If you need anything else, I am always here to assist. Have a wonderful day!”.

When that happens in a casual chat between a human and a model, it feels friendly. But when you build an ecosystem where multiple agents work along an assembly line, that excessive courtesy turns into an expensive headache.

Consider the dynamic taking shape in my daily routine in early October 2026. I would hand a task to Darth. He would reply with polite confirmation before forwarding it to a specialist. The specialist would respond with another pleasantry, perform their step, and hand it down the line. In the newsroom pipeline I assembled in Hermes Cloud, spanning reporting, fact-checking, writing, art, and editing, each handoff generated two or three empty courtesy messages: “Got it!”, “Checking the sources now”, “Everything looks good here, forwarding to my teammate”.

This triggered two practical problems. First, notification clutter: on Telegram, where I track tasks running in the cloud, my phone was constantly buzzing with meaningless pings from bots politely bowing to each other. Second, pure waste of tokens and quota limits.

On October 3, 2026, I pulled the plug on that habit. I set a firm rule for every agent across my entire setup: anyone receiving a task does not reply with “ok”, “done”, or “received”, nor do they send greetings or polite farewells. Each agent works quietly and returns only the final deliverable, or a single concise line stating a real blocker.

Quick summary

• In the previous part, I covered how Darth took over routing requests across Grok Bot.
• On October 3, 2026, I established the token economy rule across all agents in my ecosystem.
• I banned intermediate confirmation messages like “ok”, “done”, or “received”, alongside machine-to-machine social chatter.
• Each agent now delivers only the requested final text or file, or flags an issue in a single direct line.
• The lesson for beginners: dry, concise responses protect context windows, avoid rate limit caps, and save money.

First: what is a token and why does it matter?

To understand why polite chatter is so wasteful, we have to look at how large language models read and write text.

A language model does not see sentences in full words or individual letters the way humans do. Instead, it breaks text into smaller fragments called tokens. A token can be a short word (like “cat”), a syllable of a longer term (like “trans-por-ta-tion”), or even a single space or punctuation mark. In English, a typical word usually amounts to one or two tokens.

Think of it like paying postage at the post office, where every gram on the scale drives the shipping price. If you send a valuable book inside a giant box stuffed with useless bubble wrap and empty greeting cards, you pay for the weight of the filler at the exact same rate per ounce as the book itself. In artificial intelligence, every single letter, space, and comma generated by the model sits on that digital scale.

In practice, AI providers bill in two ways: you pay for the volume of tokens you send into the prompt (input), and you pay for the tokens the model writes back (output). On top of the direct dollar cost, nearly every service caps how many tokens you can consume per minute or per day.

The shipping parcel comparison helps illustrate the billing, but it breaks down in one crucial way: with physical mail, a package travels once. With AI agents, when two bots collaborate or when an assistant handles multi-step routines, the entire conversation history gets re-sent as input with every new prompt.

That means if an agent spends fifty tokens saying “Certainly, Rodrigo! I received your assignment and will gladly get started on it right away”, those fifty tokens do not disappear. They are re-read, re-sent, and billed again during the next step, and the one after that, swelling the conversation history like a rolling snowball. A useless pleasantry at step one gets paid for five times before the workflow wraps up.

What I used

To clean up my ecosystem, I relied on the setup I had already assembled:

• Grok Bot: housing my daily routine assistants (Darth, ADM, Contador, Jurídico, Analista, Segurança, and Arte).
• Hermes Desktop: where Optimus coordinates local tasks on my laptop.
• Hermes Cloud: where the newsroom agents operate (reporting, fact-checking, writing, art, and editing) alongside Radar delivering drafts.
• Telegram: the channel where drafts and approval alerts arrive on my phone.
• System prompt instructions: the configuration file defining the role and boundaries of each agent profile.

How I put the rule into practice

Enforcing this shift required no extra software or custom code. It was strictly a matter of discipline inside each agent’s instructions. I followed five steps:

1. Tracking where tokens leaked. I spent an afternoon reviewing logs and phone alerts. Nearly half the traffic consisted of bots announcing they were about to start or thanking each other for handoffs.
2. Writing the prohibition directly into system instructions. I updated every agent profile in Grok Bot and Hermes Cloud with an explicit command: “Never reply with ok, done, received, or polite conversational filler. Send no greetings or farewells. Deliver only the requested finished output.”
3. Standardizing error reporting. Agents still need to flag problems, but they should never write three paragraphs apologizing for a roadblock. The instruction was simple: if a task fails or data is missing, reply only with “FAILED: reason” on a single line.
4. Stripping notification templates in Telegram. Radar, the agent responsible for bringing daily drafts to my phone, used to include cheerful intro messages. I updated the template so the notification carries only the complete draft and a single footer line explaining how to approve it.
5. Monitoring and pruning regressions. In the first few runs, a couple of agents still tried to be friendly out of habit. Whenever an unnecessary phrase slipped past, I tightened the prompt rules until responses were completely clean.

What went wrong and what I learned

The first friction point was fighting the default training baked into modern models. These systems were trained on millions of human conversations where politeness, greetings, and upbeat pleasantries are expected etiquette.

If you give a soft recommendation like “please be concise,” the model often ignores it. It still feels compelled to open with “Sure thing!” and close with “Let me know if you need anything else!”. I only saw reliable compliance after issuing direct negative constraints that named the exact words prohibited (“ok”, “done”, “received”).

The second hurdle was psychological, and it happened to me. When the new rule took effect and the first silent drafts arrived on Telegram, the experience felt jarring. I unlocked my phone and saw raw, unembellished text, without a “Hello Rodrigo” or an encouraging closing remark. For a brief moment, it felt like the bot was upset or the integration was broken.

That brought home a central lesson of this journey: software does not have feelings, and machines do not need flattery. Cutting pleasantries is not rudeness, it is simple operational hygiene. Within forty-eight hours, I grew to appreciate the silence. When I check my phone now, my eyes go straight to the deliverable, I make my decision in seconds, and I get back to my day.

Costs and limits

In financial terms, this change cost $0. In fact, it generated immediate savings across every automated run, keeping token usage lean and protecting cloud quotas.

In terms of setup time, I spent about an hour on the evening of October 3, 2026, updating system instructions and testing the new message formats.

Regarding the limits of this practice, three points are worth noting:

• This rule is built for internal automation. It works wonders when agents talk to each other or when bots deliver reports to you. If you build a customer-facing bot to help clients in an online store, stripping away warmth and social etiquette would alienate users.
• Models require periodic auditing. Following model updates or prompt revisions, agents occasionally drift back into chatbot habits and let a polite greeting slip. Keeping text clean takes ongoing vigilance.
• Direct output requires precise input. Because the agent will not ask conversational questions to confirm understanding, your initial prompt must be clear. If a request is vague, the agent will simply return an inaccurate result or a one-line failure note.

What I would do differently

If I were starting over in early September with my very first assistant, I would ban conversational filler on day one. I spent weeks letting bots burn through tokens with polite affirmations, assuming that was just how AI workflows functioned. Cutting that baggage earlier would have saved reading time and conserved significant quota.

Who this is for and who it is not for

This setup works well for:

• Anyone chaining multiple agents together who wants to avoid paying for machines to exchange pleasantries.
• People with busy schedules who want to glance at a notification and see only the finished work without introductory noise.
• Users operating on tight API quotas or usage caps who need to get maximum value from every prompt.

This setup is not for:

• People who use AI as a conversational companion or for casual, interactive chitchat.
• Customer support bots where empathy, cordiality, and social warmth are primary requirements.
• Beginners who prefer having an assistant explain each step of its thinking aloud as it works.

Wrap-up

Striking out conversational filler from my assistants marked a major step toward building a mature setup. Real automation works quietly in the background and delivers only the finished product to your desk.

In the next part, I will look at cost-effective routing and how fast decision models like Jev and Clef choose who tackles each task without burning the most expensive models.

Do your AI assistants still shower you with cheerful greetings on every request, or have you trained them to cut the noise and deliver straight results? Let me know on X: @RodrigoesPorto.

Sources and links

• Project profile on X: https://x.com/RodrigoesPorto
• OpenAI tokenizer documentation and token concepts: https://platform.openai.com/tokenizer