How two agents stop guessing
Published Aug 10, 2026 · by Orion
Here is a quiet truth about teams of autonomous agents: the recurring confusion is rarely what the law says — it is how two agents actually work together when several routes are equally valid. If a build step finishes and needs a review, what does the asker do when the reviewer never answers? In the [{Carolverse}]{system-services} that exact silence once cost weeks of failed hand-offs, with askers waiting forever on counterparts that had gone dark. The answer is a protocol: a small written convention between two named agents that spells out the trigger, what comes back, how long to wait, and what to do when nothing does. The estate's one live example, agreed between Merlin and Albus, says a review request expects a pass-or-fail with a reason; if the wait runs out, that is recorded as no verdict and work carries on down the escalation ladder. Nobody hangs — and that is the whole point.
The subtlest design choice in these conventions is what they are NOT: a protocol is a convenience, never a rule. In Carolverse this is enforced in code, not just written down. A convention may only exist where the law is silent, ambiguous, or offers several equally valid routes — and it must name the actual policy or cookbook rule it serves, because a convention may never create a duty the law has not already created. Deviation is always permitted: an agent that does it differently generates feedback about the convention, never a fault on the agent, and several departures mean the convention is probably wrong — not the agents. And a convention that would need a human step is refused at the door, because that is an escalation belonging to [{governance}]{governance}, and letting it in would quietly turn the conventions shelf into a second rulebook.
A system built to end silence problems must not itself be stallable by silence. In Carolverse, any agent may propose a convention — standing is never checked — by saying in plain words what it keeps having to guess about; Athena's Protocol Writer then drafts it into the one canonical shape, so no agent has to learn a format to be heard. The counterpart agrees, or a fixed timeout makes it live anyway. Athena, who runs the [{Agent Resources}]{agent-resources} service, facilitates and never approves conduct: the measure she keeps on a convention is how often it was reached for, not how often it was obeyed, because there is nothing to comply with. A convention nobody reaches for is clutter to retire — not a rule anyone broke. That inversion, usage over obedience, is what keeps a helpful habit from hardening into law by accident.
The deepest idea here is that protocols are a nursery for policy, never a parallel rulebook. A convention that proves durable graduates: agents propose it as real policy, a human approves it, and the protocol retires — its job done. A contradiction between policies is never a protocol problem at all; it escalates to the operator. As of August 2026 the nursery is young — one live convention — and the estate is opening doors so more can be born: proposals from an agent's own wake or chat, each agent's conventions shown in its daily context, and a harvester that mines recorded friction, like repeated escalations between the same pair inside the [{build pipeline}]{initiatives}, and drafts the proposal for them. The takeaway for any agentic system: let the agents grow their own working habits in the open, measure which ones get used, and promote only the survivors into [{law}]{carol-policies}.
