Summary

In real multi-agent group chat usage, natural-language rule training currently works only at the session layer and can conflict with workspace rules (AGENTS.md / SOUL.md). This leads to unstable behavior across agents that use the same model.

Why this matters

A very natural product expectation is:

1. users train agent behavior through normal conversation;
2. stable rules should become durable over time;
3. newer rules should supersede older conflicting rules;
4. explicitly mentioned bots in a multi-mention message should all respond, while unmentioned bots should stay silent.

Today, session-level natural-language training can change behavior immediately, but it does not durably govern or reconcile with workspace rules. The result is role drift, session divergence, and inconsistent multi-agent group chat behavior.

Repro (real-world pattern)

Assume a Feishu/Lark group with multiple bots online.

Workspace rule (durable file rule):

• one topic, one primary responder
• target robot has exclusive priority
• non-target robots stay silent

Later, in normal conversation, the user trains a bot with natural language:

• "If you are @mentioned, you must respond, even if other robots are also @mentioned."

Then send:

• @BotA @BotB come out

Observed:

• BotA may reply
• BotB may stay silent
• or the opposite happens on another bot/session
• same model, different sessions, different behavior

I confirmed this pattern in production use with multiple agents on the same base model. The dominant factor is session history + natural-language training, not model family.

Additional observed problem

When using a weaker local model for a lightweight front-desk role, the bot may misinterpret group arbitration and start speaking on behalf of other bots (for example, "Okay, BotX, please report") even when not mentioned. Stronger cloud agents do better, but the underlying rule conflict still exists.

Expected behavior

Group chat mention semantics

• If a message explicitly @mentions one or more bots, every explicitly mentioned bot should reply once.
• Bots that are not explicitly mentioned should stay silent by default.
• If all-hands is requested, all relevant bots may respond, but only one primary responder should expand in detail.

Durable natural-language training

• Natural-language rule training should have an optional durable path.
• At minimum, there should be a governed promotion flow from session rule -> durable rule log -> workspace files / long-term memory.
• Newer durable rules should be able to supersede older conflicting rules explicitly.

Why I think this is a product gap, not just a prompt issue

The current system already has:

• workspace files injected into context;
• hooks;
• memory flush / compaction behavior;
• memory indexing.

So users naturally expect: "If I train the agent repeatedly in natural language, the system should learn durably, not just transiently in this one session."

Suggested direction

1. Add a first-class rule layer separate from SOUL.md / AGENTS.md.
2. Allow opt-in promotion of stable conversational rules to that rule layer.
3. Add conflict handling so newer promoted rules can supersede older ones.
4. In group chat, treat explicit multi-mention as a special case: all mentioned bots respond, unmentioned bots do not.

Impact

This would materially improve:

• multi-agent group usability;
• predictability of natural-language training;
• durability of agent behavior across sessions;
• reduction of session drift for same-model agents.