Skip to content
Igor Bogdanov

Architecture

Context, Reasoning, and Hierarchy in Compound LLM Agents

Deterministic context engineering provides the most reliable return per token, while distributing deliberation across an agent hierarchy can amplify hesitation and degrade decisions.

Venue
ACM proceedings paperACM CAIS 2026
Role
First author · study design · full implementation · experiments · analysis
Themes
  • Agent architecture
  • Evaluation
design axes
Three design axes of compound LLM agents: context, reasoning, and hierarchydesign space · 12 configurations · 6 models · 3,475 episodesContextwhat the agent seesraw observationsobservation historystructured network stateup to 76% betterReasoninghow much it deliberatesdirect actionquestion → critique → improvechain of thoughtHierarchyhow decisions are decomposedmonolithic agentbounded specialistscan improve returndistributed deliberationup to 3.4× worse↓ increasing cost per episode · shaded = strongest effect on each axis

Three columns labelled Context, Reasoning, and Hierarchy, each listing three levels of increasing cost. Context: raw observations, observation history, structured network state (up to 76% better). Reasoning: direct action, question–critique–improve, chain of thought. Hierarchy: monolithic agent, bounded specialists (can improve return), distributed deliberation (up to 3.4 times worse).

3,475
episodes
six models from five families, twelve configurations
283.9M
tokens consumed
fully accounted per configuration
up to 76%
lower cumulative penalty
from programmatic, structured context
up to 3.4×
worse return
from distributed deliberation across the hierarchy
1.8–2.7×
token increase
under the deliberation cascade

01

The problem

Compound agents are assembled from choices: what the model is shown, how much it is asked to reason, and whether the task is split across specialist agents. In practice these choices are made by intuition and evaluated by a single success number, so it is rarely clear which choice earned the result or what it cost.

This study isolates three axes (context, reasoning, and hierarchy) and measures each against return and token spend. The environment is again CAGE Challenge 2, a stochastic adversarial network-defence POMDP: a setting where noisy observations and a live adversary make the value of good context and the cost of hesitation visible.

The attacker, horizon, reward, and action interface were held constant. Only the agent-side design varied, so differences in return and cost can be attributed to architecture and model choice.

02

System and method

The three design axes. Context ranges from raw observations to a programmatically maintained structured network state. Reasoning ranges from direct action to question → critique → improve loops and chain of thought. Hierarchy ranges from a single monolithic agent to bounded specialists to specialists that deliberate with each other.

↑ Figure: design axes

03

Experimental design

Environment
CAGE Challenge 2 adversarial network-defence POMDP
Models
Six models from five model families
Configurations
Twelve agent configurations across the three axes
Scale
3,475 episodes · 283.9 million tokens
Primary metrics
Cumulative penalty (return) · tokens per episode · return per token
Baselines
Monolithic agent on raw observations with direct action
Cost accounting
Every model call logged with input and output tokens
Validation
Repeated seeds per configuration · cross-model replication · ACM reproducibility artifact badges

04

Main findings

  1. 01

    Structured context dominates the low-cost efficient frontier

    A programmatically maintained, structured view of network state produced the best return per token of any intervention: up to 76% lower cumulative penalty than raw observations, with no additional reasoning.

    up to 76% improvement

    Why it mattersDeterministic context engineering is cheap, reproducible, and model-agnostic. It should be exhausted before adding deliberation.

  2. 02

    Raw observations are noisy and destabilizing

    Agents fed unprocessed observations made more erratic decisions. The gap to structured context widened over the episode as errors compounded.

    Why it mattersThe observation interface is part of the agent design, not a fixed input.

  3. 03

    Bounded hierarchy can improve absolute performance

    Splitting the task across specialists with narrow, well-defined interfaces improved return for several models relative to a monolithic agent.

    Why it mattersDecomposition helps when each agent's responsibility and information are bounded.

  4. 04

    Distributed deliberation degrades every tested family

    Allowing specialists to deliberate with each other produced up to 3.4× worse return and a 1.8–2.7× token increase. The degradation is an interaction failure (hesitation propagating through the hierarchy), not simply the overhead of extra tokens.

    up to 3.4× worse · 1.8–2.7× more tokens

    Why it mattersMore agent-to-agent discussion is not a safe default. The deliberation cascade is a named failure mode to design against.

  5. 05

    Model choice changes magnitude, and occasionally sign, of architectural effects

    The same architectural change helped some models and hurt others, so conclusions drawn from a single model do not transfer automatically.

    Why it mattersArchitecture studies must replicate across model families before recommending a design.

05

What I built

The engineering behind the result. Every number above was produced by systems I designed and implemented.

  • Harness

    Hierarchical, ReAct-style agent harness over CAGE-2 with pluggable context, reasoning, and hierarchy modules.

  • Context engines

    Programmatic state-tracking layer that converts raw observations into a structured network state.

  • Agent definitions

    Configuration-driven definitions for monolithic, bounded-specialist, and deliberating-specialist topologies.

  • Model adapters

    Unified multi-provider inference with retries and token accounting for six models.

  • Experiment runner

    Parallel orchestration of 3,475 episodes with deterministic seeds and resumable runs.

  • Logging and evaluation

    Per-step trajectory logs and an analysis pipeline producing cost–performance frontiers.

  • Container setup

    Dockerized environment for the simulator, agents, and analysis.

  • Reproducibility artifact

    Archived on Zenodo. ACM reproducibility artifact badges.

06

Limitations

Contact

Building reliable AI systems requires both research and engineering.

I’m interested in research engineering, applied research, agent infrastructure, evaluation, reliability, interpretability, and research-to-production work.

Send a message

Opens your mail client with the message prefilled. Nothing is stored on this site.