Authority receipts before terminal-agent side effects

[rpelevin]

I am testing an authority-before-action receipt pattern for terminal agents.

The gap I am looking at is narrow: before an agent executes a consequential action, can the workflow prove that this exact action was authorized before execution?

For this class of workflow, the boundary seems especially relevant around shell commands, file writes, deploys, package publishes, external sends, and persistent memory writes.

Reference benchmark:
https://github.com/neurarelay/relay-action-card#agent-authority-benchmark

Dry-run command:

npm run benchmark:agent-authority -- --dry-run --json --source=github_discussion --campaign=agent_authority_week --surface=gemini_cli_terminal_authority

The benchmark is refs-only and local/dry-run: no private payloads, no downstream execution by Neura, and no claim that this project has adopted, endorsed, integrated, or been evaluated by it.

Question for maintainers/users here: is a pre-action authority receipt a useful boundary for terminal-agent workflows, or would the better insertion point be somewhere else in the execution lifecycle?

[aniruddhaadak80]

Hey there. Thanks for sharing this.

Subagents and autonomous tasks are evolving rapidly. We are working on making the coordination between multiple agents much more robust.

We appreciate your feedback and examples here.

[rpelevin]

That is the coordination boundary I would make explicit.

For terminal/subagent work, I would separate three things that often get compressed into one approval:

• parent task intent: what the user asked the agent system to pursue
• child action authority: the exact terminal action, command digest, credential scope, and policy version allowed for this step
• outcome evidence: the receipt or observation needed before the system can claim the action succeeded

The key invariant is that parent task approval should not automatically authorize a child process to read, write, execute, call a network, or claim completion. Each side-effecting action needs its own bound authority record before dispatch, and the success claim needs to join back to evidence after the fact.

That keeps multi-agent coordination robust without turning every subagent handoff into a broad permission grant.