Documentation

Agent Mode and approvals

Agent Mode lets a capable model inspect files, create artifacts, run controlled scripts, and continue across several tool steps. LlamaBoss—not the model—enforces the actual path and approval rules.

Applies to LlamaBoss v0.1.10 · Updated 2026-07-24

Turn on Agent Mode

Click the robot button for the current chat. To make it the default for new chats, enable Settings → Behavior → Start new chats with agent mode enabled.

Agent Mode works best with models trained for reliable tool calling. A small model may still chat well while struggling to produce valid tool calls.

What the agent can do

Read and search

Read files, list folders, grep text, show the working directory, and open a file in its normal Windows app.

Create and modify

Write, overwrite, edit, make folders, and delete within approved writable roots.

Documents and data

Inspect and produce CSV, XLSX, PDF, DOCX, and ZIP artifacts.

Automation

Create reviewable Python or PowerShell scripts, run approved scripts, and install a specifically named Python package.

Notes

Read or append personal notes and attached-project notes.

Web retrieval

Fetch a user-requested webpage URL when network access is appropriate.

The conversation working directory

Every conversation starts with its own workspace:

%USERPROFILE%\LlamaBoss\Workflows\chat_xxxxxxxx\Workspace

Attach a Project to add its controlled project folders. To deliberately work somewhere else, the user can run:

/cd D:\MyProject
/pwd

The model cannot secretly change the persistent working directory. A PowerShell cd inside one shell call does not persist into the next call.

Approval behavior

Not every tool shows an approval card. Clearly read-only inspections can run automatically. Commands and actions that mutate files, run broader shell automation, install packages, execute scripts, extract archives, fill PDFs, or otherwise carry meaningful risk pause for review.

ChoiceMeaning
Allow OnceRun this one requested action.
Allow AlwaysEnable one-approval mode for the current conversation so later approval-required tools can proceed without repeated cards.
DenyDo not run the tool; the model receives a denial result and can adjust.

Python package installation is always package-specific. It shows its own approval card with the exact package name even when one-approval mode is enabled.

Tool-step cap and loop guards

LlamaBoss limits how many tool results an agent can consume during one turn. View or change the cap with:

/agent_steps
/agent_steps 20

The supported range is 4–60. Repeat detection, malformed-call limits, command policy, path checks, archive protections, and output caps provide additional guardrails.

Stopping active work

Use the Stop button to cancel generation, stop an active tool sequence, or deny a pending approval before it executes. Partial assistant text may remain visible and saved when it is useful.