[Proposal] Intent-based agent addressing in threads using a System One decision model

[brennanMKE]

Motivation

In a thread with one or more agents, a human message only wakes an agent if it carries an explicit @mention. Natural follow-ups like "yes, go ahead" or "can you also check the tests?" are silently dropped, so every reply needs a fresh @.

This hits anyone working conversationally with agents in threads, and it gets worse with multiple agents. Some messages are meant for one specific agent, some for another, and some are human-to-human and shouldn't wake anyone. "Who is this message for?" is a judgment call, not something a mention-or-nothing rule can capture.

Proposed solution

Use a System One decision model to classify each human message in a thread that has agent participants:

• Input: the new message, the last N thread messages, and the participating agents (name + short role description).
• Decision: a choice over {agent_1, ..., agent_n, none}, with a confidence score.
• Action: if the top choice is an agent and confidence exceeds a threshold, treat the message as if that agent were mentioned. Otherwise do nothing.

Candidate models:

• Jev (TypeSafe AI): hosted, typed decisions, low latency and cost.
• Laya (ConvAI Innovations): open weights, runs locally. Fits self-hosted relays.
• [My Go implementation of Laya: https://github.com/brennanMKE/laya-go], which could run alongside the harness without a Python dependency.

Preferred placement is composer-side: the client classifies before sending, adds the agent's p tag, and shows "Will go to @agent" in the composer, which the user can remove. No relay or protocol change. The routing is visible and correctable, and it compiles to exactly what typing the mention produces today.

Invariants to preserve:

• Explicit @mentions always win; the classifier only fills in when no agent is mentioned.
• Agent-authored messages never trigger classification (anti-loop, matching [Bug] buzz-acp agents go deaf in threads they've joined unless re-mentioned. #2270).
• Fail closed: a low-confidence result, timeout, or error means no wake.
• Off by default, per channel or per agent, alongside "Automatically mention agents."

Alternatives considered

• Participation-based thread following ([Bug] buzz-acp agents go deaf in threads they've joined unless re-mentioned. #2270, Agent ergonomics: thread-follow after mention, and a channel-wide mention token #5867): wake agents in threads they've joined, with TTL expiry. Works well for one agent per thread, but with several agents it either wakes all of them or relies on recency, neither of which matches who the message is actually for.
• Auto-reply to the last invoked agent (Auto-Reply Toggle for Recent Agent in Threads #5096) and the existing Automatically mention agents setting: same limitation; correct only when the last agent is the intended one.
• Harness-side classification in buzz-acp: extend the [Bug] buzz-acp agents go deaf in threads they've joined unless re-mentioned. #2270 gate to queue if mentioned(agent) OR classifier(agent) >= threshold. Works without client changes, but N agents means N classifications per message unless shared, and the routing decision is invisible to the user.
• Using the agent's own LLM to decide whether to respond: accurate but slow and expensive, since every agent would spend a full LLM turn on every message.

These heuristics are a good baseline and could ship first. The classifier is a layer on top for multi-agent threads. At typical System One latency (tens to a few hundred ms), it adds little to a turn that will take an LLM seconds to answer.

Additional context

Duplicates: none found for intent classification. Closest is #2270 (thread-follow via participation tracking). Also related: #5867, #5096, #6938.

• Jev: TypeSafe AI's System One model, returns typed decisions with probabilities.
• Laya: open-source non-autoregressive System 1 decision engine from ConvAI Innovations.
• Go Laya implementation

[IRONICBo]

Composer-side looks right. Buzz already has a safer seam than inserting @Name into draft text: the address tray sends addressedAgentPubkeys through useMentionSendFlow and revalidates them before publish. I would make the model result a provisional tray entry keyed to {composerRevision, threadRoot, candidatePubkeys}, discard it on any change, and keep it out of persistent auto-pin state. That makes Remove authoritative and avoids recreating #6938 with a late result.

For evals, measure agent | none accuracy separately from whether the final p tag woke exactly one intended agent. We use that split in Jev Social too—a good route is not yet a successful action.