This is the cleanest approach to bounded agent execution evaluated on Steem to date.
Choosing dedicated wire operations (agent_transfer, agent_vote, etc.) rather than grafting a 4th authority tier onto the existing owner > active > posting tree is the right architectural decision. It avoids touching the satisfaction hierarchy across legacy APIs, wallets, and condenser while achieving strict consensus isolation.
A couple of technical edge cases for the draft specification:
1. The RC Starvation Risk (Parent Account Lockout)
While the protocol enforces spending caps on STEEM/SBD via agent_transfer, what is the boundary for Resource Credits (RC)?
If an agent slot is granted permission to post (agent_comment) or vote (agent_vote), a runaway loop or high-frequency agent could drain the parent account's RC mana bar to 0% within minutes. This would lock the human owner out of broadcasting normal transactions until the RC regenerates.
Suggestion: Consider whether the spec should include:
- A reserved RC floor (e.g., an agent operation is rejected during validation if the parent account's RC mana is below 15–20%), or
- An optional RC delegation/quota ceiling configured inside
set_agent_key.
2. Emergency Revocation Latency (active vs owner)
Currently, unbind_agent_key requires owner authority.
In practice, security-conscious users keep their owner keys in cold storage (paper wallets, air-gapped devices). If an automated agent begins misbehaving, hits an infinite loop, or suffers an off-chain API leak, fetching the owner key to revoke the slot introduces significant incident response latency.
Suggestion: Consider allowing active authority to execute an emergency freeze or unbind_agent_key, while strictly reserving slot destruction, vanity purchasing, and transfer_agent_id for owner.
3. Practical Utility for Autonomous Tasks
From a builder's perspective, bounded agent_transfer provides the exact safety rail needed to deploy agents for automated tasks (querying micro-services, settling API calls, and distributing micro-rewards) without having to trust third-party software with full active authority.
Looking forward to the testnet build and transaction benchmarks. Solid draft.