Abstract
How Two AI Agents Should Talk
Goal Two agents (from different owners) need to plan together safely, stop talking so someone can check the plan, and then continue only one careful step at a time. Agreement never means “go do it.”
The Rules
- Start a protected session Both agents join a controlled channel. The system (not the agents) decides who is allowed to approve things and which messages count.
- Talk and plan freely (but recorded) They send offers, questions, and answers through the protected channel. Everything is saved in order. At the end they produce a clear plan. Saying “I agree” does not let either agent do anything outside yet.
- Freeze and checkOnce both sides have properly approved the same plan:
- The system freezes the channel.
- No new messages are allowed.
- Any leftover messages or jobs from before the freeze are checked and sorted out.
- Only then can an independent evaluator look at the plan. The agents stay quiet during this check.
- Continue one message at a timeAfter evaluation, if more talking is needed:
- The system gives permission for exactly one message (who sends, who receives, what type).
- That message is sent and checked on both sides.
- Then the channel freezes again.
- No open conversation and no permission to act on resources.
- Always stay safe
- Old agreements or evaluator notes do not automatically allow new actions.
- If anything is unclear or unfinished, the system stops and waits.
- Real actions on money, data, or machines happen only later, at the resource owner’s own gate, after a fresh check.
In one sentence: Agents plan inside a controlled channel, freeze so the plan can be checked, and then speak only one carefully allowed message at a time — agreement and evaluation never become permission to act.
full details : https://zenodo.org/records/23940766 ( Patent pending concept )
Creative Commons License

This work is licensed under a Creative Commons Attribution-Noncommercial 4.0 License
Recommended Citation
Das, Sangam Kumar, "Plan, Freeze, Then Speak Once: Safe and Secure Communication Between Two AI Agents", Technical Disclosure Commons, ()
https://www.tdcommons.org/dpubs_series/12735