Modassembly User Guide
Types of Agents
Modassembly has 3 types of agents (that will change as the platform evolves):
- Super Agent: The most powerful kind. It can be trained by you and it can train other agents.
- External agent: An agent exposed to stakeholders outside of the company, e.g. customers. For that reason, we have to be very careful with the access that we give to it. It can’t be trained by the person speaking to it. It can only be trained by a Super Agent.
- Internal agent: Same as the Super Agent but it can’t train other agents. Meant for internal stakeholders.
How to train your agent
“To train an agent” means 2 things really:
- Show the agent how you do things.
- Help the agent map that into its own framework.
Memory
How does AI remember what you told it?
Short-term memory
Large Language Models (like Claude or ChatGPT) have a “context window”. Think of it as a bucket. This short-term memory is limited. The Super Agent’s context window size is 200,000 tokens or 150,000 words or 1 Harry Potter book. When it gets full, information starts spilling out. Think of it as a hardware limit.
Long-term memory
In Modassembly, the agent can “remember” longer than 200,000 tokens in 2 ways:
Looking back
Super/internal agents can literally search for past conversations. In that way, they can “remember” information that has fallen out of the bucket.
Knowledge base
These are articles that the agent writes down to its “workspace”. Think of it as a note for the agent.
External agents rely exclusively on knowledge bases. The knowledge base is like their manual of operations.
“WARNING: Context limit reached”
When you see this warning, it means that the agent has filled up ~90% of its 200,000 tokens and it needs to throw something out of the bucket; otherwise, it would stop completely.
While you can continue working in the same conversation session, you can see how multiple of these “compactions”, inevitably, make it forget things. The way out of it is to:
- Type:
/save. - Wait for the agent to update its workspace. Once it’s done, it will prompt you to type
/new. - Type:
/new. - It will reset its short-term memory (we now depend on the long-term memory).
Skills
A “skill” is a standard across AI Agents (Claude, OpenClaw, etc.). A skill is like a recipe of how to do something. Think of it as a knowledge base for doing things.
Where they become different from a regular knowledge base is that the agent can write code alongside the instructions. This code (written 100% by AI) is how the agent connects to APIs, etc.
The “workspace”
The workspace in Modassembly is just the place where the agent stores its knowledge base and skills.
In short
To train an agent, you need to:
- Show it your processes.
- Help it structure its memory and remember things.
- Help it write down its skills.
Tips for improving your agent
How to prompt it
Unfortunately, there are no magic prompts. Modassembly doesn’t believe that the prompts thrown around by “influencers” really work.
You can think of modern AI like a Ferrari. It can do a lot, but it’s not necessarily easy to drive.
Modassembly believes that the best way to prompt your agent is by: being curious, throwing things at it and understanding the fundamentals (memory & skills).
Improve its memory
If you notice that your agent is not recalling something, poke it to remember, poke it to search its past conversations and/or its knowledge base. Try to understand why it doesn’t remember.
Improve its quality
Once you have shown it how to do something, that will be recurring, or you have fixed a step in how it does things, prompt it to “save it as a skill”. This will improve the performance of the task because it gets hard-wired into its framework.
Review its output
Unfortunately, AI makes mistakes. Don’t take its output at face value. Do one pass, check whether it looks sound.
“My agent is getting dumb”
Your Ferrari’s motor is only as good as its bodywork. AI is the motor, Modassembly is the bodywork.
If you feel that your agent is acting clumsy, there is a legitimate chance that there is a bug in Modassembly. Reach out to Luis.
AI Credits
AI spent is calculated based on 3 factors:
- Model intelligence. As of writing this, Modassembly uses Claude (from Anthropic) exclusively. Modassembly offers the following Claude models:
Model Notes Opus 4.6 Peak intelligence, used by the Super Agent. Sonnet 4.6 Moderate intelligence, used by internal/external agents. Haiku 4.5 Lowest intelligence, usually not used, but available. Opus 5 Most intelligent, most expensive model, released on June 9, 2026. Experimental but available. - Input tokens. How long is the current conversation + your new message + new information that the agent reads as it runs.
- Output tokens. How many tokens the agent generates. Includes using “skills”.
AI credits are paid directly to Anthropic. You can find its pricing here: platform.claude.com/docs/en/about-claude/pricing.
How to save on AI credits
1. Run often: /new
The advice is to run /new everytime you switch to a completely new task. Example: you were training your agent how to generate a report and now you want to build a presentation. These are mostly two unrelated tasks. Type: /save and then: /new.
A simpler advice is to run /new once a day.
2. Keep a tidy workspace
The best way to help your agent generate less tokens is the exact same way you help it to be better at its job: keep its knowledge base and skills updated and organized.