AutoGen conversation design

Source boundary

Wu et al., AutoGen (arXiv:2308.08155v2; COLM 2024) supplies the conversation/computation/control model. Official AutoGen 0.2 documentation supplies GroupChatManager, UserProxyAgent modes, auto-reply controls, and executor configuration. Neither makes a conversation transcript complete system state or guarantees safety/correctness.

When the conversation pattern fits

Use a conversation pattern when the participants combine an LLM, tool/code result, or human input and the next reply or speaker is part of the work. For a known stage order, write the participant sequence and terminal predicate first. Use a manager-selected group only when intermediate content actually decides which role should speak next. A deterministic batch or a one-step action may need neither pattern. This is a fit decision, not a claim that conversation replaces ordinary control flow.

Choose topology, computation, and control

1. State what each participant computes: for example, an Assistant proposes code, a UserProxy executes it, and a human may supply input.
2. State who controls the next turn: direct auto-reply uses a receiving agent's generate_reply or a registered reply function; a fixed sequence names the next participant; a managed group gives GroupChatManager speaker selection and broadcast responsibility.
3. State independently what ends the exchange. A content predicate, a consecutive-auto-reply cap, and a group max_round are different controls. A successful-looking message is not an external acceptance decision.
4. Preserve received message, selected reply function/speaker, tool or external result, and terminal reason. Conversation history helps diagnosis but does not include every tool effect, external resource, hidden state, or omitted context.

Diagram scope. The direct lane abstracts the synchronous receiving-agent path with a positive max_consecutive_auto_reply (the constructed configuration below uses 4) and no custom reply function registered ahead of the built-in human/termination guard. It does not depict the source's max_consecutive_auto_reply == 0 special case, where empty human input is returned as a final user response rather than entering the normal auto-reply chain.

flowchart TB
  subgraph Direct[Direct receiver control in AutoGen 0.2.35]
    A[Assistant proposal] --> U[UserProxy receives]
    U --> DG{mode, predicate, and cap gate}
    DG -->|NEVER predicate or cap| DS[local stop: return None]
    DG -->|ALWAYS every receive| DP[prompt human]
    DG -->|TERMINATE predicate or cap| DP
    DG -->|reply permitted| DR[select auto-reply]
    DP -->|nonempty input other than exit| DU[send user response directly to peer]
    DU --> A
    DP -->|exit or empty terminal| DS
    DP -->|empty nonterminal skip| DR
    DR --> DE[LLM, code, function, or registered reply]
    DE -->|observed reply, result, or error| A
  end
  subgraph Managed[Managed group control]
    GM[GroupChatManager] --> GS[select configured speaker]
    GS --> GR[Selected group role computes reply]
    GR --> GB[broadcast reply to group]
    GB --> MG{group predicate or max round}
    MG -->|continue| GM
    MG -->|end| GX[managed-group local terminal]
  end

Direct two-agent repair trace

This is a constructed illustrative AutoGen 0.2 trace; it does not launch the framework or report a paper result. Assistant deliberately sends print(sp.sqrt(4)) to UserProxy without an import. After the receiving-agent gate permits auto-reply, the configured executor returns a concrete message such as exitcode: 1; NameError: name 'sp' is not defined. Assistant's next message identifies the changed artifact, for example import sympy as sp; print(sp.sqrt(4)), before UserProxy receives and checks it again. Record the exact tool output separately from the transcript and end with either the configured terminal predicate/cap or external acceptance.

The paper's Appendix E Table 9 is a different, successful A1 math trace: it imports sqrt, Rational from sympy, reports exitcode: 0 and 5*sqrt(42)/27, then returns TERMINATE. Use that paper trace to distinguish observed success/termination from this constructed failure repair.

An illustrative, non-executed AutoGen 0.2 configuration makes the controls visible:

def termination(message):
    content = message.get("content") if isinstance(message, dict) else None
    return isinstance(content, str) and content.rstrip().endswith("TERMINATE")
user_proxy = UserProxyAgent(
    "executor",
    is_termination_msg=termination,
    max_consecutive_auto_reply=4,
    human_input_mode="NEVER",
    code_execution_config={
        "work_dir": "./autogen-fixture-work",
        "use_docker": True,
        "timeout": 20,
        "last_n_messages": 1,
    },
)
group = GroupChat(
    agents=[planner, writer, reviewer], messages=[], max_round=6,
    speaker_selection_method="round_robin",
)
manager = GroupChatManager(group)

The displayed 4, 20, 1, and 6 are constructed local-policy values, not paper results or recommended defaults. The predicate treats absent, null, non-string, and ordinary content as non-terminal; only a string ending in TERMINATE is terminal. A real group may also configure allowed speaker transitions; the group has its own max_round and does not share the direct reply cap.

The same paper A1 math topology is shown with autonomous participation and with human_input_mode="ALWAYS". In 0.2.35, the receiving agent checks before auto-reply: NEVER does not request human input and stops locally on its termination predicate or cap; ALWAYS prompts on every receive and ignores the auto-reply cap; TERMINATE prompts at a termination condition or reply limit. A nonempty human response other than exit is returned directly as a user reply, without invoking an executor. exit, and empty input at a terminal message, resolve to a local None stop; an empty nonterminal skip follows the documented auto-reply path. Human presence is not an approval protocol.

Diagnostics and quality gates

• For a direct failure, inspect the received message, reply function invoked, executor side effect, changed repair, and terminal reason; a transcript alone is insufficient evidence of the side effect.
• For a managed group, compare configured participants/transitions and max_round with the actual selected-speaker sequence and broadcasts. An irrelevant selection is a diagnostic outcome, not a proof dynamic routing failed generally.
• Check direct reply cap and group-round limit separately. Check a human mode by observing its prompt behavior, including an unanswered prompt according to the documented mode.
• Treat an LLM critic or a TERMINATE token as advisory until an independently enforced policy controls execution or release.

Read next

• Auto-reply and custom reply functions
• Topology
• Computation and control
• Iteration
• Topology diagram
• Input/execution boundary

Bundle navigation

diagrams index.