The Deterministic Layer Anthropic’s Multi-Agent Taxonomy Doesn’t Name

Running all 5 coordination patterns in one compliance pipeline revealed a sixth

Last week, Anthropic published a taxonomy of 5 multi-agent coordination patterns (Phillips, 2026): orchestrator-subagent; agent teams; shared state; generator-verifier, and a message bus. The article offered a clear vocabulary for a field that had been naming things inconsistently. It also offered a crucial framing:

“These patterns are building blocks, not mutually exclusive choices. Production systems often combine patterns” (Phillips, 2026)

What follows is a field report on what an assembled building could look like. I am working with a team that is building an intelligent care navigation platform that analyses submissions against more than 200 regulatory controls in Australia’s disability sector and produces structured compliance assessments. The system uses all 5 of Anthropic’s patterns simultaneously. It also relies on a sixth architectural element that Anthropic’s framework does not name: a deterministic orchestration layer that runs before any AI agent, extracting clinical values, detecting contradictions, validating evidence currency, and injecting mandatory constraints into every downstream prompt. The result is a neuro-symbolic architecture: AI agents handle reasoning, deterministic rules enforce constraints. This is what makes the 5 patterns effective and safe for domains where errors matter.

Anthropic’s advice to practitioners is sound:

“We recommend starting with the simplest pattern that could work, watching where it struggles, and evolving from there” (Phillips, 2026)

We followed that advice and started with a single prompt chain (Schluntz & Zhang, 2024). Within 3 evaluation cycles, every pattern in their taxonomy had earned its place in the architecture, and the deterministic layer had emerged from the failures that revealed why patterns alone were insufficient.

Figure 1. A neuro-symbolic architecture: 5 coordination patterns coordinate a 6-stage pipeline on a deterministic foundation layer.

Five Patterns in One Pipeline

Each of Anthropic’s 5 patterns plays a distinct role in the system. Each also has a boundary: a failure mode we discovered when the pattern operated without support from the others.

Orchestrator-Subagent

Our pipeline runs 6 stages in strict sequence: document intake, evidence extraction, multi-perspective analysis, cross-perspective synthesis, compliance rating, and report generation. Each stage is an orchestrator dispatching work to specialised subagents. Stage 3 cannot begin until Stage 2 has produced validated evidence extracts. Stage 5 cannot begin until Stage 4 has resolved cross-perspective contradictions.

The boundary failure: early builds could allow Stage 5 (compliance rating) to proceed when Stage 4 (synthesis) had flagged unresolved contradictions. The rating agent, lacking explicit contradiction data, defaulted to the perspective with the strongest rhetorical framing rather than the strongest evidence. In an evaluation run, a submission with contradictory clinical evidence could receive an ADEQUATE rating when 3 of 4 perspectives had flagged insufficient evidence for the claimed outcomes.

The solution: Hard-gate dependencies between stages, where downstream processing halts until upstream outputs pass deterministic validation, will eliminate this class of error.

Agent Teams

Stage 3 deploys 4 parallel analysis perspectives: regulatory compliance, clinical evidence, participant advocacy, and accessibility. Each perspective agent receives the same extracted evidence but analyses it through a different lens. For example, the regulatory agent checks whether documentation meets legislative requirements. The clinical agent evaluates evidence quality against standardised assessment hierarchies. The advocacy agent assesses whether participant voice and choice are reflected. The accessibility agent examines whether communication and access requirements are documented.

The boundary failure: without structured output constraints, perspective agents could produce analyses of wildly different granularity. The regulatory perspective, drawing on more than 200 enumerable controls, would consistently generate 3–4x more content than the advocacy perspective. Downstream synthesis agents would interpret volume as significance, potentially weighting regulatory findings above clinical concerns.

The solution: Fixed-schema outputs with array cardinality limits (maximum 8 findings per perspective) will resolve this by forcing each perspective to prioritise.

Shared State

PostgreSQL serves as the coordination hub. Every stage writes structured outputs to the database: extracted evidence with provenance metadata, perspective analyses with confidence indicators, contradiction records, and synthesis decisions with reasoning chains.

The boundary failure: the initial designs would pass outputs between stages as prompt context. At Stage 5, the context window would contain the full output of Stages 1–4, approximately 39,000 characters. Rating agents generally show measurable recency bias, over-weighting findings from the final perspective serialised into context.

The solution: Moving to database-backed retrieval, where the rating agent queries specific evidence records rather than consuming a narrative stream, will reduce and even eliminate this bias. Early eval scores for cross-perspective consistency improved from 62% to 89% after the transition.

Generator-Verifier

Verification operates at 2 scales. Micro-verification validates every AI output against a JSON schema before it enters shared state. If the clinical evidence agent returns a finding without a required evidence_level field, the output is rejected and the agent is re-prompted with explicit schema instructions. Macro-verification runs after synthesis: a dedicated Safety Verifier agent reviews the complete assessment for internal consistency, checking whether the compliance rating is supported by the evidence cited and whether participant risk factors identified in early stages are addressed in the final report.

The boundary failure: micro-verification alone could miss semantic errors. An agent could return a perfectly schema-valid output that misclassified evidence quality. Schema validation simply cannot catch clinical misclassification.

The solution: The macro Safety Verifier, operating on the assembled output with access to the original evidence extracts, will catch these errors by comparing claimed evidence levels against the deterministic extractions from Stage 2.

Message Bus

The system uses Server-Sent Events for the user interface, streaming stage completions and progress indicators to the front end. A message bus handles inter-agent communication, enabling agent-to-agent negotiation during Stage 4 (synthesis) and supporting deployment of specialist agents without modifying the orchestration logic.

The boundary failure: without a message bus, inter-agent communication would run through shared state alone. Contradiction resolution during synthesis would take perspectives in a fixed order, with the last perspective processed having outsized influence on the resolution framing.

The solution: The message bus enables simultaneous resolution during Stage 4, removing ordering effects. It also decouples agent deployment from orchestration logic, allowing specialist agents to be added to the pipeline without restructuring existing stages.

Table 1. Each pattern serves a specific function and has a specific failure mode. The resolution column shows how inter-pattern dependencies and the deterministic layer address each boundary.

The Interactions Between Patterns Are Where Complexity Lives

Individual pattern failures are relatively simple to diagnose. The harder problems emerge from interactions between patterns, where a behaviour that is correct within one pattern creates failures in another.

Cross-Stage Framing Drift

Stage 2 (evidence extraction) feeds Stage 3 (multi-perspective analysis). When a submission contains strong positive evidence for one domain, the extraction agent’s framing of that evidence carries rhetorical momentum into the perspective agents. A submission with excellent clinical documentation but poor safeguarding practices could receive ADEQUATE ratings across all 4 perspectives. The clinical evidence would be strong enough that perspective agents anchor on it, even the advocacy perspective, which should flag the absence of documented consent processes.

The interaction: The orchestrator-subagent pattern (Stage 2 feeding Stage 3) interacts with the agent teams pattern (4 parallel perspectives). Strong extraction framing in Stage 2 would bias all 4 team members in Stage 3, defeating the purpose of multiple perspectives.

The solution: A deterministic normalisation step strips rhetorical framing from evidence extracts and presents findings as structured key-value pairs. Perspective agents receive data, not narrative.

Cross-Perspective Contradiction

Stage 3 produces 4 perspective analyses. Stage 4 synthesises them. The clinical perspective could rate a support plan as having insufficient evidence (no peer-reviewed basis for the intervention approach), while the regulatory perspective rates the same plan as compliant (all required documentation is present). Both assessments would be correct within their frame. Without explicit rules for handling this category of disagreement, the synthesis agent could produce a compromise assessment that satisfies neither perspective and obscures a genuine clinical concern.

The interaction: The agent teams pattern (4 independent perspectives) creates contradictions that the orchestrator-subagent pattern (sequential synthesis) cannot resolve without explicit rules. Legitimate disagreements between perspectives would collapse into compromise assessments that obscure clinical concerns.

The solution: A deterministic contradiction detection step between Stage 3 and Stage 4 flags opposing assessments of the same evidence item, classifies each contradiction by type (clinical-regulatory, clinical-advocacy, regulatory-accessibility), and injects a resolution protocol into the synthesis agent’s prompt.

Timeout Hierarchy Mismatch

Stage 3 (multi-perspective analysis) dispatches 4 parallel agents, each making multiple LLM calls, potentially with extended thinking enabled. A stage-level timeout set to match a single agent's expected completion time, say 90 seconds, would create a mismatch. Complex submissions could push individual perspective agents to 150–180 seconds, well within their individual budgets but far beyond the stage timeout. The system would retry, hit the same limit, and return a partial assessment missing perspectives.

The interaction: The orchestrator-subagent pattern (stage-level timeouts) conflicts with the agent teams pattern (parallel fan-out). A timeout calibrated for a single agent breaks when the same stage dispatches 4 agents in parallel.

The solution: The stage timeout is computed as architecture rather than set as configuration: stage_timeout = max(individual_agent_timeout) × 1.5 + orchestration_overhead. For Stage 3, with 4 parallel agents each allocated up to 210 seconds, the stage timeout accommodates worst-case completion. This computation is deterministic and adjusts automatically if agent timeouts change.

The Missing Deterministic Layer

Anthropic’s multi-agent coordination patterns describe how AI agents interact with each other. They do not describe what happens before AI agents run. In any domain where system outputs carry real consequences and errors compound through downstream stages, what happens before AI agents run is the most consequential architectural decision.

In our pipeline, what runs before the agents is the deterministic layer: a set of rule-based processing steps that execute at defined points, always before the AI agents they feed.

Clinical Value Extraction

Before any AI agent analyses a submission, a deterministic parser extracts structured clinical values: standardised assessment scores, evidence quality levels, timeframes, and regulatory identifiers. The parser uses pattern matching, table extraction, and domain-specific ontologies.

Feldman et al. (2025) demonstrated that generalist LLMs confabulate clinical values at rates between 7% and 15% when extracting structured data from medical documents, even when explicitly instructed to quote verbatim. A 10% confabulation rate on evidence levels in a compliance system means roughly 1 in 10 assessments could cite high-quality evidence where only self-report exists. The deterministic parser eliminates this risk category entirely. Downstream AI agents receive pre-extracted clinical values as injected context, marked as ground truth.

Contradiction Detection

Before the synthesis agent (Stage 4) runs, a deterministic comparator analyses the structured outputs from all 4 perspective agents. It identifies assessment conflicts, evidence disputes, and threshold disagreements. Each contradiction is recorded in shared state with a typed classification and a resolution protocol. The synthesis agent receives these contradictions as mandatory context.

Evidence Currency and Hierarchy Validation

Clinical evidence has temporal validity. The deterministic layer checks evidence dates against domain-specific currency windows and flags expired evidence. It also validates evidence hierarchy claims: if an agent classifies evidence at a certain quality level, the deterministic layer checks whether the source metadata supports that classification.

Mandatory Constraint Injection

Every AI agent prompt in the pipeline contains a block of deterministic constraints, injected programmatically, that cannot be overridden by the agent’s own reasoning. Initial eval results show that agents comply with injected constraints at 97.3% fidelity, compared to 71.8% when the same information is provided as narrative context within the main prompt body. The structured injection format matters as much as the content.

Rating Rubrics with Hard Gates

The compliance rating agent operates under a rubric that references upstream deterministic outputs by identifier. Hard gates make certain rating outcomes impossible when deterministic checks have failed, removing the AI agent’s discretion on questions where discretion introduces risk.

Delvecchio et al. (2025) describe this architectural pattern as neuro-symbolic AI: the combination of neural network capabilities with symbolic rule systems. Their survey across multiple domains supports the finding that hybrid architectures consistently outperform pure neural approaches on tasks requiring both reasoning flexibility and output reliability. Our architecture arrived at this pattern empirically, through eval cycles, before encountering the formal taxonomy.

Four Things We Learned by Running Evals

Anthropic’s evals guidance recommends treating evaluation as a continuous engineering practice rather than a quality gate (Anthropic, 2026). 4 findings from our eval cycles shaped the architecture.

1. Rubrics Beat Model Upgrades

We ran the same eval suite against a larger model (Opus 4.6) and a smaller model (Sonnet 4.6). The expectation was that the larger model would produce consistently better compliance ratings. It did not.

Opus 4.6 achieved 78% rubric alignment. Sonnet 4.6 achieved 64%. Both showed the same pattern: inconsistent application of evidence hierarchy rules. After introducing deterministic evidence hierarchy validation and hard-gate rubrics, Opus 4.6 achieved 96% and Sonnet 4.6 achieved 93%. The 14-point gap collapsed to 3 points.

Table 2. Rubric alignment before and after the deterministic layer. The layer reduced the performance gap between models from 14 points to 3 points.

For compliance-critical outputs, investment in deterministic constraint architecture yields higher returns than investment in more capable models.

2. Output Constraints Prevent Verbosity Drift

Without output constraints, perspective agents produce outputs averaging 39,000 characters combined. With structured schemas and array cardinality limits, combined output drops to 12,400 characters with no loss of substantive content in human review.

The reduction has downstream effects beyond token cost. Agents processing 12,400 characters of structured findings show 27% improvement in cross-reference accuracy compared to processing 39,000 characters of narrative output. The constraint forces perspective agents to prioritise, and the prioritised outputs are easier for downstream agents to process accurately.

3. Macro-Verification Catches Inter-Agent Bias

The Safety Verifier flags an average of 2.3 issues per assessment. The most common flag (41%) is “rating not supported by cited evidence.” The second most common (28%) is cross-perspective bias: the synthesis agent giving disproportionate weight to one perspective.

Without macro-verification, eval assessments contain unsupported ratings at a rate of 23%. With macro-verification and automated re-assessment on flagged items, that figure drops to 4%.

4. Timeouts Are Architecture

Without dynamic timeout computation, timeout mismatches account for approximately 11% of incomplete assessments in eval runs. Dynamic timeout computation based on stage composition eliminates timeout-related failures entirely.

Operational parameters like timeouts, retry limits, and context window budgets are architectural decisions in multi-agent systems. A timeout that works for a single-agent call breaks when the same call fans out to 4 parallel agents. These parameters must be computed from the system’s structure, not set as static defaults.

What the Stack Looks Like

Anthropic gave practitioners a vocabulary of 5 patterns and explicitly stated they are building blocks meant to be combined. The vocabulary is accurate. The patterns are necessary. Every one of them addresses a real coordination problem in multi-agent systems!

But patterns describe interactions between AI agents, and they do not describe the substrate those agents stand on. In any domain where system outputs carry real consequences, where errors compound through downstream stages and are difficult to detect or reverse, that substrate cannot be probabilistic. Production multi-agent systems for high-stakes domains are neuro-symbolic pattern stacks, and the bottom layer of the stack is deterministic.

References

Anthropic. (2026, January 9). Demystifying evals for AI agents. https://www.anthropic.com/engineering/demystifying-evals-for-ai-agents

Delvecchio, G. P., Stilo, G., & Melo, G. D. (2025). Neuro-symbolic AI: A task-directed survey. In Proceedings of the International Joint Conference on Artificial Intelligence. https://arxiv.org/abs/2603.03177

Feldman, M. J., et al. (2025). Dedicated AI expert system vs generative AI for clinical assessment extraction. JAMA Network Open, 8(5). https://doi.org/10.1001/jamanetworkopen.2025.12994

Phillips, C. (2026, April 10). Multi-agent coordination patterns. Anthropic. https://claude.com/blog/multi-agent-coordination-patterns

Schluntz, E., & Zhang, B. (2024, December 19). Building effective agents. Anthropic. https://www.anthropic.com/research/building-effective-agents