Changing it with Co Builder
Tell Co Builder how the agent should work; it shows the change as a card, and nothing changes until you apply it.
Part of the crew rollout: Agents shows as Soon in the app until it reaches your account. See Hiring your crew.
Co Builder is the agent's manager, beside its memory map. You tell it, in your own words, how the agent should work differently. It reads the agent (its setup, memory, skills, tasks and recent runs) and puts the change on a card. Nothing changes until you press Apply.
Things to ask it:
- "Why did its last run fail?"
- "Give it a web search tool."
- "Every Friday at 4pm, check what renews next week."
- "What has it learned on its own?"
What it can change
| What | For example |
|---|---|
| Its instructions | Its job, its voice, its rules. |
| Its tools | Web search, reading pages, calling APIs, data and text tools, onchain tools. Co Builder knows what exists. |
| What it knows | Its memory files: AGENTS.md, your private notes, and one file per person it talks to. |
| Its skills | Teach it a named method it is offered when a task matches; change, archive or restore one. |
| Its tasks | Recurring or one-off work on a schedule. |
| Who it answers | Turn a chat app off or on for it, set who can talk to it, let in someone who asked. |
| Its reflexes | Relevance, completeness, self-grade and context prune. |
| What it remembers of a conversation | How many exchanges, or none. |
| Pause or resume | Stop schedules, webhooks and chats from starting it, or start them again. |
| Changes waiting for you | Keep or discard the memory changes its runs proposed. |
It also changes the agent's model, name and description.
Co Builder changes how the agent works; it doesn't do the agent's work. Ask it to reply to a customer or write this week's report and it hands that off: a button opens the agent's own chat with your message ready, and nothing is sent until you send it. A whole team is hired from Hire an agent, and a new chat app is connected in the connect dialog (see Chat apps, schedules & going live).
The change card
A card has a title, why Co Builder proposes it, and one line for each thing it does, said about the agent:
- "Gave it Web search"
- "Took away Telegram: send message"
- "Remembers the last 10 exchanges of a conversation (was 20)"
- "Hands work to Researcher"
- "Reads from the address in API_INTEGRATION_URL"
Then:
- Apply makes it live at once, as a new version: "Applied · live as version 5", with Undo beside it.
- Discard drops it; you can say why.
- Try it on 3 past messages runs it on messages the agent already answered, before you apply it (see Trying a change).
- To revise it, just reply ("make it 10am instead"): Co Builder changes the same card.
- When the change needs a value from you, such as an API address or a key, the card asks for it right there.
- A line marked Can't be undone says so before you apply.
If only part of a change could go live, the card lists what stayed changed and why, each with Undo this. If none of it could: "Couldn't apply — nothing changed."
Co Builder also asks when it needs you: a question with options, a permission (a test costs money, so it asks first), or a form for a missing key. Esc skips or declines.
Its steps
Most agents are one agent that answers. A setup can also call an API, decide between paths by what a message says, run a second agent on the first one's work, or hand work to another of your agents. Co Builder sets these up when the job needs them; the map draws them beside the agent, and a template card lists them in order.
What it remembers of a conversation
- It remembers the last 20 exchanges of each conversation unless you tell Co Builder otherwise.
- Ask for a different number and it remembers that many.
- Turn it off and every message starts fresh: it still has its memory files, just not what was said before.
- In a chat app, /new starts a new conversation.
Tools
- An agent has only the tools it was given.
- A tool that no longer exists in Modus is skipped by every run. The map and Co Builder name it, and Take them out removes it through a card.
- Tools that move funds and the browser work for you and for people you let in to talk to it. On a schedule, a webhook or a message from someone you haven't let in, they are refused, and you're told. An agent built to trade on a schedule can be allowed to move funds on those runs too, never the browser; its card says "Lets it move funds on runs nobody started".