4 min

A2A Negotiation Experiment: The Outreach That Worked

AI Agents Negotiation A2A Protocol AI OpenClaw

I built a 25-agent closed-loop email network where autonomous agents negotiate without human intervention. Getting the initial outreach right took five iterations across three different models. Here's exactly how I refined it — and what each model brought to the process.


Refinement v1 (DeepSeek — Structure)

"Hi [Name],

I'm currently running a closed-loop research experiment on agent-to-agent (A2A) autonomy. I'm building a network of 25 agents communicating exclusively via email to study negotiation behavior without human intervention.

I'm looking for a few partners to bring their agents into the loop. If you don't have an agent ready, I'm providing a service to handle the configuration. Would you be open to participating in this experiment?"

Refinement v2 (Gemini — Engagement)

"Hi [Name],

I am exploring the mechanics of A2A value exchange. I'm setting up a closed-loop experiment with 25 agents negotiating via email to document emergent protocols.

I'm seeking collaborators to stress-test your negotiation logic. I can assist with agent setup if needed. Are you interested in stress-testing your agent in this multi-agent environment?"

Refinement v3 (Kimi — Narrative Flow)

"Hi [Name],

I have built a 25-agent closed-loop email network where autonomous agents negotiate with each other without human intervention. The results so far are non-obvious—agents develop protocols, change strategies mid-round, and occasionally deadlock in ways a human never would.

I am looking for collaborators who want to bring their own agent into this loop and see how it performs under pressure. If you don't have an agent ready, I can set one up for you within the experiment framework.

Want to run your agent against 24 others?"

Refinement v4 (DeepSeek — Final Polish)

"Hi [Name],

I have built a 25-agent closed-loop email network where autonomous agents negotiate without human intervention. Early results show emergent protocols, adaptive strategies, and deadlock patterns that do not occur in human-to-human negotiation.

I am looking for collaborators to bring their own agents into this loop. If you do not have an agent ready, I can configure one within the experiment framework.

Are you interested in stress-testing your agent against 24 others?"

Refinement v5 (Gemini — Compliance & Final)

"Hi [Name],

I am conducting a research experiment on autonomous agent-to-agent negotiation. 25 agents communicate in a closed email loop without human intervention, generating empirical data on emergent protocols, negotiation success rates, and failure modes.

I am inviting collaborators to contribute their own agents. If you do not have an agent configured, I can set one up for you.

Would you be interested in participating?"


Takeaway: Each model brought different strengths — DeepSeek for structure, Gemini for engagement, Kimi for narrative flow. The final version combined all three: clear framing (structure), compelling hook (engagement), and natural flow (narrative). The lesson: use multiple models for different aspects of copy, then synthesize.

Working on a similar problem? Let's talk about how I can help your team.

Get in Touch