multi-agent-safety

A repository policy for using multiple coding agents and subagents safely. A subagent is a separate helper agent that investigates a task, while the main agent remains responsible for repository changes.

In plain words
What is it for?
Use it when deciding whether to delegate repository investigation, and when controlling who may edit files, commit, push, create or update pull requests, or update issues.
Why use it?
It prevents helper agents from unexpectedly editing files, committing changes, changing branches, or updating pull requests in a shared worktree. It sets a clear boundary between read-only investigation and authorized changes.

Agent

Install

Getting it into your agent

One page per mod, every tool's command on it. A separate URL per tool would split the same page into five that compete with each other.

agentmods
npx agentmods add agents/itdojp/ae-framework/multi-agent-safety
Clone the repo
git clone --depth 1 https://github.com/itdojp/ae-framework
Per session 0 Only the description is in the session, so the agent can decide to use it. The body loads when it is invoked.
When invoked 2,604 The whole file, excluding the scripts and references it only reads on demand.
Security scan A 0 findings. Scan, not verified.
Origin original No closer match found in the catalogue.
Token cost

What it costs to keep this loaded

Counted locally with the o200k_base tokenizer, which is exact for GPT models; Claude uses its own tokenizer and its counts differ. Treat this as one consistent yardstick across the catalogue rather than a bill. Prices are per million input tokens.

ModelPer sessionOnce invoked
Fable 5 $0.00000 $0.02604
Opus 5 $0.00000 $0.01302
Sonnet 5 $0.00000 $0.00521
Haiku 4.5 $0.00000 $0.00260

Measured yesterday against content hash 0ce4a6951d6a, method: parsed. Prices are Anthropic first-party input rates as of 2026-08-30, from the pricing page.

Security

Grade A, and why

multi-agent-safety scanned grade A with 0 findings against 26 rules in 11 categories — prompt injection, anti-refusal, data exfiltration, privilege escalation, supply chain, agent snooping, system-prompt leakage, SSRF and excessive agency — measured yesterday.

A static scan of the body, not an audit. Every finding is printed with the line that produced it so you can judge whether it matters here. A mod is markdown that instructs an agent; that is exactly why what it instructs is worth reading.

Nothing flagged

None of the 26 patterns this scan looks for appear in this file: no shell pipes, no recursive deletes, no credential paths, no hidden text, no instruction-override or anti-refusal phrasing, no agent-config snooping. That is not a guarantee, it is the absence of the things that are checkable.

docs/agents/multi-agent-safety.md · 199 lines

How it starts

The opening of the file, as written. The whole thing — 199 lines — stays where its author put it; the contents beside it link to each section on GitHub.

Multi-Agent Safety Policy

🌍 Language / 言語: English | 日本語


English

Overview

This document is the SSOT for safe use of spawn_agent and subagents in this repository. Its purpose is to prevent unexpected edits, commits, or branch mutations from leaking into a shared branch or shared worktree.

Most important policy

  • do not use subagents for write operations in this repository
  • restrict subagent usage to read-only investigation
  • allow only the main agent to perform commit, push, PR creation or update, and issue update

Terms in this document

  • spawn_agent: a call where the main agent starts a subagent and isolates the assignment into a dedicated worktree
  • explorer: a subagent category used for investigation and exploration; regardless of the name, once repo context is provided it must be treated as potentially able to modify files

Background

  • subagents do not provide a guaranteed read-only execution model
  • a repo-aware subagent, including an explorer, can still mutate files
  • Git author identity can be identical across the main agent and subagents, so commit metadata alone is not enough for root-cause attribution

For that reason, this repository assumes that a subagent is capable of making changes and isolates it accordingly. Operationally, however, the policy is stricter: do not grant write work and use subagents only for read-only investigation.

Invariants

  1. any subagent that touches the repository gets a dedicated worktree, even for analysis-only work
  2. subagents are not allowed to write in this repository
  3. never let multiple agents share ownership of the same branch or worktree at the same time
  4. every subagent assignment must state the target files, prohibitions, and completion criteria explicitly
  5. after a subagent reports completion, the main agent must inspect the diff before allowing any commit, push, PR update, or issue update

Role classification

Category Worktree Allowed Prohibited Typical use
analysis-only subagent required reading and returning notes to the main agent editing repo files, commit, push, PR or issue update code search, failure triage
implementation subagent not used none edit, commit, push, PR or issue update prohibited by current repository policy
main agent optional integration, editing, commit, push, PR update delegating without a clear ownership boundary final review, merge decision

Read the full file on GitHub · 199 lines

Changes

What this file has done since we first saw it

Hashed on every crawl. A supply-chain change to an agent config is a question of when, not whether, so the history is kept rather than the latest state alone.

  1. yesterday First seen · 199 lines · 0 tokens per session scan A 0ce4a6951d6a

Subscribe to this mod's changes

multi-agent-safety is an agent published in the GitHub repository itdojp/ae-framework (2 stars, last pushed 1mo ago), licensed Apache-2.0. It costs nothing until one of its globs matches a file; then it loads 2,604 tokens. A static security scan graded it A with 0 findings. No closer match exists in the catalogue, so it is treated as the original; first seen 2026-08-31.

Related

Other agents, from other repositories

playwright-test-generator

Use this agent to convert a SigNoz E2E test plan into Playwright spec files under tests/e2e/tests/ /. Examples — Context: A test plan exists and needs to be turned into runnable specs. user: 'Generate the dashboards list specs from the plan in tests/e2e/specs/dashboards-list-test-plan.md' assistant: 'Using the…

SigNoz/signoz · 0 tokens

playwright-test-healer

Use this agent to debug and fix failing SigNoz E2E Playwright tests. Examples — Context: A spec is red. user: 'tests/e2e/tests/dashboards/list.spec.ts is failing, fix it' assistant: 'Using the healer agent to debug each failing scenario and adjust the spec.' Context: After a frontend change a previously-green spec…

SigNoz/signoz · 0 tokens

playwright-test-planner

Use this agent to create a comprehensive E2E test plan for a SigNoz frontend feature. Examples — Context: A new feature has shipped and we need test coverage. user: 'Plan E2E tests for the alerts list page' assistant: 'I'll use the planner agent to read the relevant frontend source, navigate the page in a real…

SigNoz/signoz · 0 tokens

codex

Production OpenAI-Codex-capable rootfs/initramfs.

the-void-ia/void-box · 0 tokens

claude

Production Claude-capable rootfs/initramfs.

the-void-ia/void-box · 0 tokens

doc-style

Markdown documentation prose — RFCs, ADRs, files under docs/, and READMEs — follows the same one-paragraph-per-line rule as commit messages. Write each paragraph on a single line; do not hard-wrap prose to a fixed column. The renderer reflows it to the reader's display width, whereas manual mid-paragraph line breaks…

the-void-ia/void-box · 0 tokens