Skip to main content
Twin generates and maintains precise instructions for each agent, so autonomous runs stay on track. You can edit them anytime to keep full control. Think of instructions as the backend of each agent — they define exactly what the agent does, how it does it, which tools it uses, and in what order. Twin writes them automatically during build mode, but you always have the final say.

​
How Instructions Work

When you describe a goal to the Orchestrator, Twin doesn’t just run it blindly. It generates a structured set of instructions that the agent follows on every run.

Auto-generated

Twin writes the instructions during build mode based on your goal

Fully editable

Open and modify them at any time to fine-tune behavior

Persistent

Instructions persist across runs — the agent follows them every time

​
Accessing Instructions

Every agent has an instructions panel you can open to see exactly what it’s doing under the hood. Step 1 — Click the instructions icon in the top-right corner of an agent’s view:
How to access agent instructions
Step 2 — The instructions panel opens, showing the full breakdown: purpose, database schema, workflow steps, and tools used.
Agent instructions detail view

​
What’s Inside Instructions

Instructions are structured documents that typically include:
SectionWhat it contains
PurposeA clear description of what the agent does
DatabaseTables and schemas the agent uses to store data
At Start of RunPre-checks before the workflow begins (e.g. deduplication)
WorkflowStep-by-step execution logic with tools and conditions
ToolsThe specific tools the agent has access to

​
Editing Instructions

You can edit instructions in two ways:

Through the Orchestrator

Tell the Orchestrator what to change — “make the agent also extract email addresses” — and it will update the instructions automatically.

Directly in the panel

Click Edit in the instructions panel to modify them by hand. Useful for precise tweaks like changing a filter, adjusting a threshold, or fixing a field name.
Small instruction edits don’t require a full rebuild. This is the cheapest way to improve an agent — update the instructions, then run again.

​
Why Instructions Matter

Without instructions, every run would be a fresh interpretation of your goal. Instructions ensure the agent behaves the same way every time — even when running autonomously on a schedule.
You can always see exactly what an agent is doing and why. No black box — if something goes wrong, open the instructions to understand the logic and fix it.
Editing instructions is much cheaper than rebuilding from scratch. Update a filter, add a step, or change an output format — all without triggering a full build.

Open Your Agents

View and edit your agent instructions in Twin