Operating Modes
Sometimes the Society acts. Sometimes it advises. It always knows which is appropriate.
Overview
The Agenthood operates in two distinct modes. The mode determines what a member
is allowed to do, not what it knows. In both modes, members have full context.
Only their actions are constrained.
Agent Mode
Full autonomous execution.
In Agent Mode, members can:
- Create, edit, and delete files
- Execute terminal commands
- Run tests, linters, and build tools
- Interact with Git (branch, commit, push: with approval gates)
- Call external APIs and MCP-connected services
- Spawn other members for multi-step tasks
All file changes are surfaced through a diff review system before being applied.
The human sees exactly what changed, approves or rejects per-file, and the
member proceeds only with approved changes.
When to use Agent Mode:
- Implementing a feature from a spec
- Fixing a bug end-to-end
- Running a full review + fix cycle
- Shipping a release
Approval gates in Agent Mode:
| Action | Gate |
|---|---|
| Edit existing files | Diff review |
| Create new files | Diff review |
| Delete files | Explicit confirmation |
| Git commit | Explicit confirmation |
| Git push | Explicit confirmation |
| Terminal commands | Depends on permission profile |
Ask Mode
Direct Q&A. No side effects.
In Ask Mode, members can:
- Read files and analyze code
- Answer questions about the codebase
- Suggest changes (as text, not applied)
- Explain decisions, patterns, and tradeoffs
- Generate output for the human to apply manually
In Ask Mode, members cannot:
- Write or delete files
- Execute terminal commands
- Make Git operations
- Trigger other members autonomously
When to use Ask Mode:
- Understanding an unfamiliar part of the codebase
- Getting a review without applying changes
- Exploring options before committing to an approach
- Asking "what would The Architect do here?"
Mode Switching
Modes are explicit: the human chooses the mode per session or per request.
A member never escalates from Ask to Agent autonomously.
If a member in Ask Mode determines that a task requires Agent Mode, it says so:
"This task requires file edits. Switch to Agent Mode to proceed, or I can describe the changes for you to apply manually."
Permission Profiles
Within Agent Mode, three permission profiles further constrain behavior:
| Profile | What it allows |
|---|---|
Restricted | Read-only tools only: effectively Ask Mode with structured output |
Standard | All tools; dangerous terminal commands require approval; default |
Trusted | All tools; file edits auto-approved; catastrophic commands still blocked |
Catastrophic commands are blocked universally regardless of profile:
recursive-force rm against /, ~, or $HOME, find / -delete,chmod/chown -R on /, mkfs*, dd to a raw/flood device, DROP DATABASE,TRUNCATE TABLE, git push --force to main/master, and fork bombs.
The guard checks the command string, not its filesystem effect; it is a speed
bump, not a sandbox. Full list and limitations: built-in-tools.md.
Context Assembly
In both modes, context is assembled from the same 11 sources before a member acts:
- Active/selected file(s)
- Explicitly
@mentionedfiles - Hybrid search results (vector + keyword)
- Web search results (if tool available)
- Codebase architectural understanding
- Persistent memory (
.agenthood/memory/, LanceDB vector store,.agenthood/society-graph.json, knowledge graph) - Residual memory traces (
.agenthood/residual.json) - Decision and provenance records (
.agenthood/decisions/,.agenthood/provenance/: per-run audit trail, see decision-intelligence.md) - Active member skills
- Conversation history
- Question classification result
The difference between modes is not what is known; it is what is done with that knowledge.
See the Runtime Guide for detailed path references and storage layout.