Simulated: no real agent, service, or token standard is used. The root key stays in this page. This is a teaching model, not a production token format.
Agent chat
Token chain: what the receiving tool sees
Live check with crypto.subtle. Each tag is HMAC-SHA256 of the earlier tag and the caveat.
Try this
- Press Step twice. Watch the bars shrink and the chat show each hop.
- Press Start over. Hand off to B, but tick email.send. Read the reason it fails.
- Hand off to B and C. Press Tamper, then Verify chain. See which hop fails first.
Pain point
A second agent can end up with more power than the first agent had.
Problem statement
A sub-agent that reuses its parent's key has all of the parent's power. How might we hand work on safely?
What sparked it
Mozilla's W3C WebExtensions notes (issue 1041) say no browser primitive lets a user scope or withdraw an agent's access.
Why this approach
Each handoff may add limits, never remove one. The root key checks the whole chain.
Hypothesis tested
We believed that a hash of the earlier tag lets each hop only add limits. If it holds, a widened token fails.
How to try it
- Open the Prototype tab. Press Play.
- Press Start over. Hand off from A to B, and ask for more than A has.
- Hand off from B to C. Press Verify chain.
- Press Tamper. Verify again.
What this does not do
- It uses no real agent. All agents are simulated.
- The root key stays in the page.
- The token format is a teaching model, not for production.
Notes
- Who would pay: agent framework vendors and platform teams.
- First customer: a team with multi-agent workflows in production.
- Other options: a policy server, or signed receipts.
- Weak spot: someone must manage each hop's key.
- Feedback question: does each refusal reason say what to change?
Traces
| Step | Prompt | Model | Tool | Date |
|---|
Cost
| Step | Model | Input tokens | Output tokens | Cost (USD) |
|---|---|---|---|---|
| Total | ||||
Measured: the author runs used 186,075 tokens in total. Estimated: the split into input and output tokens (90 percent input, 10 percent output) and the dollar cost, at $2 input and $10 output per million tokens for Sonnet 5.5 (about $0.52). The run reported one total only. The lead session's planning, research, and publishing are not included. The prompts in the Traces section are listed in the order they were given.