Agent safety, permissions, and reversibility
Tool scope and permissions
Scope manifests and tool permissions
- A Tool Registry With Tiers: Who May Call What
- Per-Caller Caps Layered on Top of Tool Tiers
- Per-Caller Caps Over Tool Tiers: Weights and a Floor
- Permission Tests That Catch a Widened Subagent Tool
- Subagent Tool Scoping: Limiting What a Delegated Task Can Touch
- Tier Claims in the Token Versus a Lookup per Call
- A Clearance Check at Execution: Wrapping Every Tool Call
- A Scope Manifest per Agent: Verbs, Resources and Owners
- All of Tool scope and permissions
Prompt injection defense
Prompt injection defense and sanitization
- A Prompt Injection Test Suite You Can Run in CI
- A Quarantine Release Checklist: Proving a Chunk Is Safe
- Detecting Poisoned Chunks Before They Reach the Prompt
- Grading Injection Defenses: A Scorecard per Attack Class
- Injection Tests for Tool Results
- Memory Poisoning: How Bad Facts Get In and How TTLs Limit Harm
- Prompt Injection Through User Data: A Boundary and a Test Suite
- Tool Permissions as a Second Line of Injection Defense
- All of Prompt injection defense
Dry runs and idempotency
Dry runs, confirmation, and idempotency
- A Dry-Run Endpoint for Reversals: Testing the Undo First
- Idempotency Keys for Agent Writes Behind a Human Gate
- Two-Phase Writes From the Agent's Side: Dry Run, Then Commit
- A Commit Token That Carries the Plan: Fields and a TTL
- A Confirmation Token the Model Cannot Forge
- A Read-Only Rehearsal of a Write Tool Before It Fires
- Choosing Between JSON Schema and Pydantic for Agent Contracts
- Contract Tests for a Dry-Run Endpoint
- All of Dry runs and idempotency
Reversibility and undo
Reversibility and undo of agent actions
- A Reversal Plan per Write Type: Undoing an Approved Action
- A Compensating Call That Fails: When the Undo Needs Its Own Undo
- A Fake Ledger That Rejects an Unbalanced Reversal
- A Reversal Plan per Agent Write Type
- Classifying Agent Actions by Reversibility
- Compensating Writes When the Commit Succeeds
- Giving the Agent the Undo Tool: Rules for Reversing a Write
- Proving a Reversal Ran Once: A Ledger and Its Checks
- All of Reversibility and undo
Kill switches and denials