Conductor-Style Planner

• Conductor-style planner is a modular LLM agent that dynamically constructs and executes workflows by interleaving introspection, tool calls, and state updates.
• It decomposes tasks via context specialization, enhancing precision and reducing hallucinations through adaptive and transparent shared-state management.
• Empirical studies show it achieves faster convergence and higher accuracy in data discovery and preparation compared to static pipelines.

A conductor-style planner is a modular, LLM-driven agentic orchestration mechanism for dynamic workflow management, particularly suited to ill-specified or evolving goal states. It provides adaptive control over multiple agentic and retrieval-capable subsystems by interleaving introspective reasoning, tool invocation, shared-state mutation, and user interaction. The paradigm is central to systems where the specification of intent or the solution pathway is not known a priori, but iteratively constructed or refined in the course of interaction, as exemplified by the Pneuma-Seeker platform for data discovery and intent alignment (Balaka et al., 7 Jan 2026).

1. Concept and Definition

A conductor-style planner is an LLM agent that dynamically constructs and executes complex workflows by selecting, at each iteration, from a menu of action types rather than following a fixed pipeline. The core cycles of an LLM conductor involve:

• Internal reasoning (with access to shared state and user feedback)
• Tool calls to external modules or subsystems
• State modification of a shared canonical representation (e.g., a relational schema plus queries)
• User-facing communication for clarifications or progress reporting

Decisions are made on-the-fly based on evidence of gaps between the current artifact (such as synthesis state) and the active user intent, without static execution graphs or rigid templates. In the Pneuma-Seeker architecture, the conductor orchestrates IR, data preparation, and user feedback resolution in a context-specialized, modular fashion (Balaka et al., 7 Jan 2026).

2. Architectural Position and Action Space

Within a modular, agentic system, the conductor agent occupies the top-level planning role. It maintains both a view of the shared state—typically represented as a tuple $(T, Q)$, with $T$ a set of tables and $Q$ a sequence of queries—and a contextual model of user intent, which evolves through iterative feedback.

The conductor's available actions (illustrative from Pneuma-Seeker) are:

Action Type	Description
Internal Reasoning	ReAct-style introspection on state and incoming feedback
Tool Call	Invocation of IR, Materializer, SQL Executor modules
State Modification	Editing $T, Q$ to reflect updated schema or logic
User Interaction	Messaging to clarify, request information, or report

The next action is chosen with a budgeted introspection cap (e.g., after five internal cycles, force user-facing communication) to prevent runaway internal deliberation (Balaka et al., 7 Jan 2026).

3. Shared-State Convergence Mechanism

A defining principle is the maintenance of a shared, explicit workspace or state, commonly a relational schema and associated query list $(T, Q)$. The conductor and the user interact by proposing and refining this representation at each turn. Feedback from the user or from downstream tool results is used by the conductor to mutate shared state, driving toward a state that sufficiently satisfies the latent information need. Convergence is formally defined as the stabilization of state $(T, Q)$ and explicit user approval:

$$\text{converge} \iff \exists K \quad S_K = S_{K+1} \land \text{user approves } S_K$$

where $S_k = (T_k, Q_k)$ at iteration $k$. This explicit externalization of planning state enables transparency, traceability, and rapid correction of misalignment (such as missing or erroneous columns) early in the workflow (Balaka et al., 7 Jan 2026).

4. Context Specialization and Tool Integration

Unlike monolithic LLM agents that operate with all context and tool invocation logic in a single prompt, the conductor-style planner decomposes responsibilities via context specialization. Each component LLM—Conductor, IR System, Materializer—receives only the state and history necessary for its subtask. By reducing prompt breadth and length, this mitigates both hallucination and performance degradation due to context-window limits.

Tool calls within this paradigm are standardized: queries or commands to IR systems, data materializers, or execution backends (such as SQL). Each tool’s interface and expected return payload are manifest and routable by the conductor, which interprets the results in the context of the current planning state (Balaka et al., 7 Jan 2026).

5. Application in Semi-Automatic Data Discovery and Preparation

In the Pneuma-Seeker system, the conductor-style planner underpins iterative workflows for bringing semi-structured or poorly-specified information needs to fulfillment. The core sequence proceeds as:

1. User posits a vague, high-level question.
2. The conductor introspects on current $T, Q$ and user feedback to identify gaps.
3. It invokes retrievers to discover candidate tables, then materializes joins or transformations.
4. The conductor refines or edits schemas/queries in response to both automated tool output and iterative user clarifications.
5. Actions iterate until state convergence and user approval.

This flexible, non-linear planning loop outperforms static pipelines and traditional RAG workflows on tasks requiring significant evolution of schema and query designs, as shown by convergence rates and accuracy in simulated user studies (Balaka et al., 7 Jan 2026).

6. Evaluation and Empirical Findings

Empirical assessments of conductor-style planners have been performed using LLM-simulated users on data discovery and preparation scenarios:

• Convergence rate (questions resolved within 15 turns): ≈83% (Archeology), ≈90% (Environment)
• Median turns to convergence: ~4–6
• Outperformance of table-retriever and RAG baselines in fit-for-purpose answer generation
• Token cost and latency are higher than pure retrieval, but accuracy and intent alignment are significantly improved
• The conductor’s iterative planning captures organizational memory, as each schema refinement, join strategy, and user clarification is logged for future reuse (Balaka et al., 7 Jan 2026).

7. Broader Significance and Limitations

The conductor-style planner is well-suited to domains where information needs cannot be captured by static pipelines or fixed retrieval, and evolution of shared state is critical. Its central affordance is adaptive, context-sensitive orchestration of agentic and retrieval-oriented submodules under explicit user alignment loops.

Limitations include higher per-turn latency and token cost, reliance on effective tool interfaces, and inherent dependence on the quality of state representations and user feedback loops. There is ongoing research in scaling such planners to multi-table or more general multi-objective synthesis workflows, as well as extending context specialization to encompass richer agent ecosystems (Balaka et al., 7 Jan 2026).

A plausible implication is that conductor-style planning constitutes a foundational architecture for next-generation interactive systems that require not only retrieval or generative capacity, but also explicit, interpretable intent alignment and organizational memory capture across evolving cycles of user and system co-design.