Outborn Agent Platform · Step-by-step guide
Agents
Create reusable agents that combine instructions, models, governed sources, tools, skills, memory, channels, automations, and approval boundaries.
Before you begin
- Choose the owning Space and a narrow agent purpose.
- Prepare approved sources or tools only if the agent needs them.
Follow the workflow
- 01
Define the agent
Set its purpose, instructions, expected response, and boundaries for actions it should not take.
- 02
Select model and context
Choose an available runtime source and attach only the permitted knowledge, data, skills, and connections needed for the job.
- 03
Set access and review rules
Check who can use the agent and whether connected writes require approval.
- 04
Try representative work
Run a normal request and an out-of-scope request. Inspect the answer, tool activity, and any approval proposal before wider use.
The agent handles its intended task and stays inside the configured access and action boundary.
Check runtime availability, source grants, agent access, and the specific failed tool or approval state.
Applied guidance
Best practices for agents
Recommended
- Write a narrow purpose, expected output, and action boundary in the agent instructions.
- Trial normal, ambiguous, and out-of-scope requests before wider use.
Avoid
- Do not attach every available connection merely because it exists in the Space.
How it fits together
01
Creating agents
Create an agent for a repeatable role or workflow rather than putting every instruction into each conversation.
02
Instructions
Define the agent's purpose, boundaries, expected outputs, and important operating rules.
03
Runtime and model sources
Choose one or more runtime/model sources and keep each source's account, model, priority, and usage boundary explicit.
04
Sources and priority
Attach approved knowledge, data, applications, and other sources. Priority should describe preference, not bypass permissions.
05
Knowledge
Attach reusable Knowledge Bases for semantic retrieval and source citation.
06
Data
Attach governed Data Collections for structured analysis and business context.
07
Applications and tools
Grant the agent only the connected tools it needs. Consequential writes can remain approval-gated.
08
Skills
Skills package reusable instructions and tool behavior that can be attached to agents according to permissions.
09
Memory
Configure durable memory separately from conversation context, company knowledge, and connected data.
10
Channels
Publish or connect an agent to supported communication channels only after channel permissions and identity behavior are understood.
11
Automations
Scheduled and conditional work should use the same agent sources, permissions, and approval boundaries as interactive work.
12
Publishing
Publish only with the intended audience and access model. Public distribution should not expose private organization resources.
13
Permissions
Agent access never grants more authority than its organization, Space, source, connection, or user permissions allow.
14
Human approval
Use approval gates for consequential external actions so the proposed change remains visible before execution.
Related documentation