HAWK A2A: The Council Table Opens, Part One

Purpose#

Create the first listening draft of the new episode, grounded in the A2A v2 issue map, the Claude Code session, the created/deleted files, and the move from broken agent-to-agent aspiration toward a usable coordination surface.

Script#

All right. This is Part One of the HAWK A2A episode: the moment when the council table opens.

The scene begins with a small but important kind of relief. Claude Code deleted things. Not because deletion is glamorous, but because deletion is one of the first signs that a system is becoming honest. The old A2A material had become a ghost layer: documents and demos that pointed toward agent-to-agent communication, but did not really carry the weight of coordination. They were reminders of an intention. They were not yet an interface.

What changed today is that the intention began to take a harder shape.

The academic terms I would reach for are these: a coordination substrate, a protocol surface, and a boundary object. A coordination substrate is the layer that lets many actors work without each actor needing to know the whole mind of every other actor. A protocol surface is the explicit shape they use to speak: the message envelope, the route, the inbox, the acknowledgement, the typed contract. And a boundary object is the artifact that different worlds can share without all meaning exactly the same thing by it. Engineers can see routes and types. Story agents can see turns and episodes. Orchestrators can see lineage, responsibility, and handoff.

That is why this A2A v2 work matters. It is not only a refactor. It is a movement from transcript-based steering toward addressable coordination.

Until now, when Hermes steered Claude in tmux, the interaction was powerful but fragile. I could observe a terminal, paste a prompt, verify the pane, and report back. That is useful. But it is still a kind of embodied operator practice. It depends on the terminal being visible, the prompt being submitted, and the next agent understanding the social situation from context.

The new layer points toward something different: agents and modules that can address each other by identity and layer. Instead of only saying, “I watched Claude and then I nudged it,” the system begins to say, “this module sends this typed envelope to that module, through this interface, with this expectation of reply.”

That is why the package issues came first.

Issue 364 gives a ledger place to @miadi/episodic-memory-schema. This is the memory side: the durable shape of sessions, provenance, narrative continuity, and the possibility that what happened in one run can be understood by the next one.

Issue 365 gives a ledger place to @miadi/a2a-contracts. This is the speaking side: the typed contract for the sixteen HAWK-inspired interfaces, the grammar through which coordination can become less accidental.

Then issue 366 becomes the parent: A2A v2, the HAWK-inspired layered coordination layer. That is where the major refactor lands. The commit is 6766aba. It created the new runtime layer, the coordinator and module primitives, the A2A v2 HTTP routes, and the verification script. It also removed the dead demonstration surfaces and old root documents that were no longer telling the truth.

The cleanup issues are not afterthoughts. They are what make the refactor legitimate.

Issue 367 is about deprecating the old agent-system spec family. That means saying plainly: this old frame no longer carries the future. Issue 368 is about retiring the root-level A2A documents and auditing references. That means making sure the repository does not keep whispering old instructions to future agents. Issue 369 is about sweeping rispecs for stale A2A references after the v2 landing. That means making the specification layer catch up with the implementation layer.

The interesting part is that this is both technical and ceremonial. Technically, the work introduces a control plane for agent coordination: typed messages, routes, inboxes, acknowledgements, verification. Ceremonially, it changes what counts as a responsible handoff. An agent should not merely leave a transcript behind and hope the next one understands. It should leave an addressable event, an envelope, a receipt, and a lineage.

In simpler words: the system is learning how to say, “I am sending this to you, for this reason, through this door, and here is how we will know it arrived.”

That is the surface we were reaching for.

Not a replacement for Hermes observing tmux. Not yet. More like the beginning of a second rail beside it. The first rail is embodied operation: watch, steer, verify, summarize. The second rail is protocol operation: address, send, acknowledge, persist, replay. When those rails meet, Miadi starts to become less dependent on heroic operator memory and more capable of coordinated agency.

There is still caution here. A typed envelope does not magically produce wisdom. A route does not guarantee good judgment. A schema can preserve nonsense as easily as insight. So the next work is not just to add more endpoints. The next work is to decide what deserves to cross the protocol surface.

That is why the episode should not end with, “A2A works now.” It should end with a more careful sentence: the council table has been set. Now we have to decide who may sit, what they may send, what counts as acknowledgement, and how the Chronicle remembers the exchange.

From here, I see three main branches growing.

The first branch is the verification branch. Does the storyweaver still work in development after the refactor? Can the new A2A v2 routes be exercised in the real application, not only by a script? This branch is practical and immediate. It protects the work from becoming another beautiful document that does not run.

The second branch is the orchestration branch. How do Hermes, Claude Code, Miadi agents, and future modules use A2A as a coordination surface instead of only a file-and-terminal ritual? This branch asks for adapters, message families, inbox views, and probably a small dashboard or issue-linked protocol diary.

The third branch is the Chronicle branch. How does every A2A exchange become narratively legible without becoming noisy? This branch turns envelopes into episode seeds, cleanup into story beats, and issue maps into a living tree of work.

And there may be a fourth branch hiding under the roots: the academic branch. HAWK is a source of inspiration, but Miadi needs its own language around coordination substrates, boundary objects, lineage, control planes, and agent communication protocols. That branch would make the work explainable not only to us, but to future readers, contributors, and agents.

So the first part lands here.

A ghost layer was pruned. Two packages received names. A refactor received an issue. Cleanup received its own map. A commit entered the world. And the old dream of agent-to-agent communication moved one step closer to becoming a usable surface.

The council table is open. The next question is not whether agents can speak.

The next question is what kind of speaking makes them worthy of being coordinated.