Agent orchestration occupies a multi-dimensional design space

Type: kb/types/note.md · Tags: computational-model

Agent orchestration architectures are often discussed as if they sit on a single progression: raw chat loop, better chat loop, symbolic scheduler, versioned infrastructure. That framing collapses several independent design choices into one axis. The result is unstable taxonomies where every new system requires "adding another point" to a list that was never one-dimensional to begin with. The better picture is a design space with separable dimensions, so new systems can occupy new combinations without rewriting the taxonomy.

Scheduler placement

Where the scheduler lives — in the LLM conversation itself (LLM-mediated schedulers are a degraded variant of the clean model) or in symbolic code / an external runtime (bounded-context orchestration model). The distinction is where transition-relevant state and progression live: inside bounded stochastic context, or in explicit symbolic state outside a call. The latter can meet the normal form's conditions; it is a placement dimension, not a ranking of architectures. RLM is a boundary case — the scheduler lives on an exact substrate outside chat even when the model authors some of that scheduler code.

Decomposition-policy artifact

What kind of system-definition artifact carries the policy for choosing decompositions. This is separate from scheduler placement. The scheduler executes a decomposition; the policy selects which decomposition to run, when to split further, and what scaffold language counts as available.

The policy can occupy different representational forms (how it is encoded and consumed) and storage substrates:

  • Natural-language — human-authored or trace-extracted heuristics in prompts, playbooks, notes, or skills.
  • Symbolic — executable schedulers, generated orchestrator code, task graphs, routing programs, or restricted combinator libraries.
  • Distributed-parametric — a learned policy in model weights, adapters, or controller state, trained to choose decompositions inside a codified scaffold.

Representational form does not say how the policy was acquired. A natural-language policy may be written upfront, generalized from execution traces, or generated by an agent after failure analysis. In the authors' LongCoT-mini case study, an initial RLM(GPT-5.2) run scored 50.6%; trace inspection found brute-force timeouts and failures to check sub-agent answers; after Claude Code wrote mistake-avoidance tips from the trajectories and restarted the run, the source reports 65.6% overall. This is a trace-derived natural-language intervention, but the post does not analyze the updated traces and therefore does not establish that behavior shifted to checked graph decomposition. The intervention is not yet trained policy or durable verified code, but it is more structured than ad hoc prompting.

RLM illustrates another mixed point: the model authors symbolic orchestrator code, but the result is ephemeral rather than a durable repo artifact. λ-RLM retains RLM's prompt-as-environment REPL while replacing open-ended model-authored control code with a typed library of deterministic, pre-verified combinators. A deterministic planner instantiates a fixed recursive chain, while neural inference is bounded and localized to task selection, leaf work, and explicitly specified synthesis steps. This improves analyzability and predictable execution. The retained complete-system description ends at response return and does not specify the post-return lifecycle of REPL state or the execution plan. Those extracts therefore cannot establish whether λ-RLM shares standard RLM's per-task persistence boundary. The Mismanaged Geniuses Hypothesis proposes first defining the space of decompositions available to a language model and then allocating training effort to models that choose correct decompositions within a scaffold. Treating that as a codified decomposition language with a distributed-parametric selection policy is this note's placement, not the source's terminology or a source-described codify-then-relax migration. Representational form, acquisition path, storage substrate, and persistence can therefore vary independently even when the visible scheduler shape looks similar.

Persistence horizon

How long orchestration knowledge persists — per-step only (conversational loops), per-session (Slate's episodes retained across a task), or cross-session (versioned scheduler logic, reusable workflow artifacts). The cited RLM account externalises the loop and describes within-run persistence; classifying its post-return lifecycle requires separate evidence. Versioned orchestration pushes the same pattern into durable infrastructure; Slate occupies an intermediate point where compressed products persist within the session without becoming long-term artifacts.

Coordination form

How bounded contexts coordinate — direct prompt assembly, back-and-forth conversation, prompt refinement, shared-state / blackboard, or context cloning / forking. Conversation vs prompt refinement establishes the narrower distinction between conversational continuation and prompt refinement, including their hybrid and their placement under symbolic-scheduler and LLM-caller architectures. The larger list is this note's proposed synthesis; the cited comparison does not establish that every listed form is independent or combines with every scheduler placement.

Coordination guarantee

What failure mode the architecture prevents when coordination happens. The form of coordination does not tell you whether the architecture is safe — a shared substrate can exist with or without the guarantee matched to its composition mode. Agent orchestration needs coordination guarantees, not just coordination channels develops this in detail. It separates three failure families over shared semantic substrates:

  • Isolation / scoping — prevents contamination
  • Consistency / ownership / visibility — prevents inconsistency in shared state
  • Adjudication / verification / voting — prevents error amplification in combined outputs

It also identifies a fourth, governance family: liability firebreaks or authority refresh prevent an accountability vacuum in delegation chains.

Boundary-return artifact

What comes back from a bounded execution — raw natural-language output, structured extraction, compressed episode / branch summary, or symbolic code / control program. The return artifact determines what the next stage can do cheaply: a compressed episode supports orchestration differently from a raw transcript; code supports exact re-execution differently from a natural-language summary.

Why this matters

The multi-dimensional framing explains why single-axis taxonomies keep breaking. Slate adds a new combination because it changes persistence horizon and boundary-return artifact simultaneously. Forking changes coordination form without changing scheduler placement. Treating both as "extra points on one line" obscures what each system actually varies.

The value is not taxonomy for its own sake but sharper questions:

  • Which dimensions are structural and likely to survive scaling?
  • Which dimensions interact, and which vary independently?
  • Where do current systems cluster, and which regions are unexplored?

This should stay an open map, not a closed classification. The current dimensions are salient, not final.

Relevant Notes: