Agentic automation

Agentic automation is useful when a task cannot be described as a fixed list of RPA blocks. Khadim gives the agent tools, context, and guardrails so it can decide which actions to take at runtime.

How a run works

A Khadim run follows the same loop across CLI, desktop, web, and SDK surfaces:

1. You provide a task, prompt, automation, or trigger.
2. Khadim builds the system prompt from the active mode, tools, project instructions, memory, and runtime settings.
3. The model replies with text, tool calls, or a question for the user.
4. Khadim executes tool calls, streams events, and returns observations to the model.
5. The loop continues until the agent emits done or an unrecoverable error.

Good first tasks

Start with tasks that have clear evidence of completion:

• Summarize a repository.
• Find and fix one failing test.
• Convert a recurring shell workflow into a script.
• Extract data from a local file and produce a report.
• Inspect a screen and explain what action is needed.

Use project instructions

Add AGENTS.md files to teach Khadim stable project conventions:

# Agent instructions

- Use `bun test` for TypeScript changes.
- Keep migrations separate from application changes.
- Ask before running long integration tests.

Prefer durable facts and conventions. Avoid one-off instructions that only apply to the current task.

Choose the right surface

Use the surface that matches the task:

Surface	Use when
CLI	You are working in a repository or scripting a one-off run
Desktop	The task needs local UI, screen, browser, or credential context
Web	A team needs monitoring, managed agents, or shared run history
SDK	Your application needs to embed Khadim as a capability
Docker	The run needs isolation, reproducibility, or server deployment

Keep automation observable

Prefer prompts that tell Khadim what evidence to collect:

Fix the failing auth tests. Run the smallest relevant test command before and
after the change, then summarize the files changed and the test result.

This gives the agent a concrete completion target and gives you a useful audit trail.