Agent instructions
Learn how durable instructions define an Agent's role, working rules, and expected behavior.
Agent instructions define how an Agent should approach its role across repeated work.
They are durable guidance for the Agent, so the user does not need to restate the same rules in every Chat.
What belongs in Agent instructions
Use Agent instructions for behavior that should stay true across future work.
Good candidates include:
Role
What the Agent is responsible for.
Working rules
The recurring rules and constraints it should follow.
Terminology
Company-specific language the Agent should use consistently.
Output expectations
How the Agent should structure or present recurring work.
Instructions vs one-off Chat requests
Keep durable behavior in Agent instructions.
Keep temporary directions in the current Chat.
| Put it in Agent instructions | Keep it in Chat |
|---|---|
| “Always cite the source of customer facts.” | “Summarize these 3 tickets.” |
| “Use our internal lifecycle terminology.” | “Make this answer shorter.” |
| “Confirm before taking an outward-facing action.” | “Focus on Acme for this request.” |
| “Prioritize open opportunities over closed ones.” | “Research these 5 accounts today.” |
If a direction should still matter next week, it is a strong candidate for the Agent’s durable instructions.
Instructions vs context
Instructions tell the Agent how to behave.
Context tells the Agent what it should know.
For example:
- “Always separate confirmed facts from assumptions” is an instruction.
- “Our enterprise tier is called Scale” is context.
The two can work together, but keeping them conceptually separate makes the Agent easier to maintain.
Learn about Agents and context →
Instructions vs Skills
A Skill stores a reusable method for a particular kind of task.
Agent instructions define the broader operating behavior of the Agent.
For example:
Recruiting Agent instructions
Defines the Agent’s role, tone, decision boundaries, and recurring working rules.
Candidate screening Skill
Defines a reusable process for evaluating a candidate against a role.
If the method could be useful to another Agent, it probably belongs in a Skill.
Write instructions that leave room for judgment
Useful Agent instructions describe outcomes and boundaries without turning every task into a rigid script.
Prefer guidance such as:
- “Return the strongest evidence first.”
- “Ask before taking an outward-facing action.”
- “Use live CRM data when the Connection is available.”
- “Flag missing evidence instead of guessing.”
Avoid bloated instructions that prescribe every click, search, or tool call unless that detail is genuinely required.
Connections can have their own operating guidance
Some Connections can include provider-specific guidance about how they should be used.
That guidance can complement the Agent’s own instructions.
For example, a browser-based Connection may carry service-specific operating notes while the Agent instructions define the Agent’s broader role.
The live Connection configuration remains the source of truth for provider-specific behavior.
Learn about Agent Connections →
Editing Agent instructions
Agent instructions are durable context.
When editing existing instructions, Giga should read the current version first and amend it rather than accidentally wiping useful guidance.
Internally, the context tools may treat this durable Agent guidance as the Agent’s brief.
Read before replacing
The underlying context update can replace the stored content. Existing instructions should be reviewed before a meaningful update so useful guidance is preserved.
Meaningful changes should be confirmed with the user before they are saved.
Keep instructions maintainable
A useful Agent instruction set is usually easier to maintain when it has a small number of clear sections.
For example:
Role
Research target accounts for the Sales team.
Priorities
Prioritize current company facts and evidence from approved Connections.
Working rules
- Separate confirmed facts from inference.
- Link to sources when available.
- Ask before making outward-facing changes.
Output
Return a concise account summary, useful signals, and open questions.
This gives the Agent a stable operating model without freezing every task into a fixed procedure.
A simple test
Before adding something to Agent instructions, ask:
“Should this still guide the Agent across future work?”
If yes, it belongs in durable instructions.
If it only matters for the current task, keep it in Chat.
If it is a reusable task method, consider a Skill.
If it is a multi-step process, consider a Workflow.