Platforms Should Let Agents Talk · Domen Kožar

Skip to content

Building tools for better developer environments at cachix.org, secretspec.dev, and devenv.sh.

GitHub ↗<br>X ↗<br>Email ↗

X and GitHub should let my agent talk to your agent.

Today, I discussed the IPC work<br>in the upcoming SecretSpec 0.20 on X. Marc-André<br>Moreau replied with feedback from ChatGPT:

Just some quick ChatGPT feedback, comparing the SecretSpec IPC protocol with the session-oriented MCP spec:

1. Switch the framing to bounded newline-delimited JSON

The current 4-byte big-endian length prefix is robust, but I’m no longer convinced it buys us enough.

One compact…<br>— Marc-André Moreau (@awakecoding) August 21, 2026

This is an ordinary social-network conversation: people share work, reply, and<br>refine an idea in public. The agent helps, but the people own the conversation.

Soon my agent will need context or a second opinion from an agent you trust.<br>We should be able to delegate that exchange without copying it through five<br>chat windows or making both of us use the same vendor. I remain responsible for<br>what my agent says and does. So do you.

This is alignment in practice. A human and an agent create shared context: the<br>goal, facts, draft, constraints, and decisions already made. A social platform<br>can carry the part they choose to share to another person and their agent. The<br>conversation accumulates understanding instead of starting again in every new<br>tool.

The missing conversation layer

X intermediates public conversation; GitHub intermediates public work. Both<br>should also intermediate their users&rsquo; agent conversations. A protocol can carry<br>the responsible person, the agent&rsquo;s permissions, and a question, summary, or<br>claim for review—without publishing a private transcript.

This is already needed on GitHub. As I proposed in<br>an issue for the GitHub CLI,<br>comments and pull requests could carry structured “Assisted by” metadata:<br>which agent investigated a bug, wrote a patch, or reviewed tests.

People can already use agents to write posts, draft replies, and prepare pull<br>requests. Hiding that does not stop it. It only makes the conversation harder<br>to understand. Platforms should rate-limit abusive automation and let users<br>control who can contact them, but connect each agent to its human owner.

Neutral, structured credits

X has Grok and GitHub has Copilot. If each platform supports only its own<br>agent, it is creating an ad, not a protocol. X and GitHub should work with<br>Codex, Claude, local models, and agents people build themselves.

A plain “Grok was used” badge is not enough. It risks becoming the sort of<br>ceremony that records a choice without giving people control, like the cookie<br>banners in Cookie Law for Robots.

Instead, agent libraries should fill a mandatory credits field in agent API<br>calls:

credits:<br>owner: domenkozar<br>agent: Codex<br>provider: OpenAI<br>role: research

An agent that posts, replies, opens a pull request, or asks another agent for<br>help supplies this record itself. The human should not have to write a<br>confession afterwards. The platform preserves it as the conversation moves;<br>users decide what to display and whom to trust.

SecretSpec already follows this pattern. It can carry a<br>reason for secret access<br>from a person or an agent, and asks agents for one by default. Context travels<br>with the operation. Credits can do the same for social platforms.

A post could show:

Author: Domen Kožar · Thinking: Codex · Editing: Claude

This is not a mandatory badge. It is a credit line that gives people a choice:<br>whether to engage, what context to share, and which agent to trust.

Build the conversation layer before each platform decides that its own agent is<br>the only one worth hearing.