Planner–Executor–Reviewer Pipeline

• The PER pipeline is a structured agent architecture that separates planning, execution, and reviewing to enhance accuracy, traceability, and robustness.
• It employs a staged workflow where a planner decomposes goals, an executor implements plans using tools, and a reviewer validates outcomes for error detection.
• Practical applications include multi-tool QA, visual reasoning, code testing, and SDLC, yielding measurable performance improvements and verifiable results.

A Planner–Executor–Reviewer (PER) pipeline is a structured agentic architecture wherein distinct components—specialized for planning, execution, and verification—are strictly decoupled, allowing for robust, transparent, and verifiable completion of complex tasks across application domains. This architectural pattern, dominant in both academic and industrial agentic AI, enforces a staged workflow: the planner decomposes high-level goals into executable sub-plans, the executor materializes these plans through tool interactions or reasoning, and the reviewer validates each outcome, yielding iterative feedback or termination. Rigorous stage separation, structured handoffs, and orchestrated loops (with or without adversarial model families) are systematically shown to improve accuracy, stability, traceability, and verifiability in long-horizon, open-ended reasoning and automation tasks.

1. Formal Definitions and Canonical Roles

The PER pipeline comprises three principal agent roles, typically instantiated by LLMs or neuro-symbolic systems:

• Planner: Receives a high-level goal $G$ and (optionally) contextual memory $C$ and reviewer feedback $f$, outputting a plan $\Pi$. Conceptually:

$\Pi = P(G; C, f)$

The plan is often an ordered sequence or a graph of sub-questions, sub-tasks, or natural-language prompts, and may be encoded as a structured object (e.g., JSON, script, code, API call graph).

• Executor: Consumes the plan $\Pi$ and contextual tool interfaces $T$, generating the candidate artifact $A$:

$A = E(\Pi; T)$

Here, $A$ can comprise tool outputs, code artifacts, intermediate results, or reasoning chains. In multi-modal settings, the executor may invoke vision models or specialized utilities.

• Reviewer: Given artifact $C$0 and a verification mechanism $C$1, computes executable feedback $C$2, which can be pass/fail, structured diffs $C$3, or more elaborate audit objects:

$C$4

The reviewer may inspect syntactic/semantic correctness, run tests, cross-check claims, or compute rubric-based trace scores.

The pipeline composes as follows:

$C$5

until a termination predicate $C$6 (e.g., $C$7) is satisfied (Apostolou et al., 14 May 2026).

2. Data Flow, Interaction Loop, and Orchestration

The PER pipeline is orchestrated as a loop, where planning, execution, and review iterate in a closed feedback cycle. The system may be managed by an explicit orchestrator, which routes artifacts and feedback among agents according to strict schemas:

$\Pi = P(G; C, f)$9

A typical runtime progression is:

1. Planner: Decomposes $C$8 (with $C$9, $f$0) $f$1 $f$2
2. Executor: Executes $f$3 (with $f$4) $f$5 $f$6
3. Reviewer: Audits $f$7 (with $f$8) $f$9 $\Pi$0 (feedback or diff)

This staged workflow can be enriched by:

• Context-saving strategies and external memory in executors to mitigate context overflow (e.g., partial previews and persistent storage for large tool outputs) (Molinari et al., 3 Dec 2025).
• Structured handoff schemas (e.g., strict logging of prompt, intermediate answers, error origin) for full traceability and blame assignment (Barrak, 8 Oct 2025).
• Protocols for adversarial cross-family review (e.g., executor and reviewer from distinct LLM families) (Yang et al., 4 May 2026).

3. Verification, Guardrails, and Feedback Loops

The Reviewer establishes verifiability by synthesizing executable feedback:

• Stage-wise Veracity: The reviewer applies a set of verification procedures, such as unit test harnesses, static analyzers, output validators, or ensemble module cross-checks. Each check is typically a Boolean predicate or a structured rubric (Apostolou et al., 14 May 2026, Xu et al., 9 Jun 2025).
• Refinement Loop: Reviewers return either pass (success) or fail (with actionable $\Pi$1), triggering local or global re-planning. Formal schema guardrails constrain the executor action space and ensure output adherence to predefined schemas:

$\Pi$2

where $\Pi$3 is the language defined by output grammar $\Pi$4.

• Executable Reward Modeling: In reinforcement learning settings (e.g., TraceLift), reviewers combine solution correctness (verifier pass/fail) with intrinsic reasoning trace quality (reward models $\Pi$5) to tightly couple intermediate plan utility to downstream execution uplift (Han et al., 5 May 2026).

A table contrasting PER stages is below:

Role	Input	Output
Planner	$\Pi$6, $\Pi$7, $\Pi$8	Plan $\Pi$9
Executor	Plan $\Pi = P(G; C, f)$0, toolset $\Pi = P(G; C, f)$1	Artifact $\Pi = P(G; C, f)$2
Reviewer	Artifact $\Pi = P(G; C, f)$3, verification $\Pi = P(G; C, f)$4	Feedback $\Pi = P(G; C, f)$5 (pass/$\Pi = P(G; C, f)$6)

4. Practical Instantiations and Performance Impact

The PER pipeline is instantiated in various domains:

• Enterprise Multi-tool QA: In RP-ReAct, the Reasoner-Planner Agent generates abstract sub-steps, the Proxy-Execution Agent executes them in a ReAct loop, and the RPA’s reviewer monitors each result, invoking local or global replans as needed (Molinari et al., 3 Dec 2025). Empirically, mean accuracy on hard tasks improved from 0.13 (monolithic executor) to 0.19 (PER, std 0.05), a 40% relative gain.
• Visual Reasoning: VLAgent employs a Planner that emits scripts, an Executor integrating neuro-symbolic modules, and Reviewer modules that parse, repair, and ensemble-verify outputs. Each layer adds measurable accuracy: e.g., on GQA, SS-Parser repair increases accuracy from 58.9% to 61.2%, with ensemble verification further to 61.9% (Xu et al., 9 Jun 2025).
• Code and Math: In TraceLift, PER is used to train a planner whose reasoning traces are executed by a frozen executor and scored by both a reasoning reward model and an executor-based uplift. This yields substantial micro-average accuracy increases (e.g., +4.4 points code, +22.7 points math over two-stage baseline) (Han et al., 5 May 2026).
• Industrial SDLC Applications: A comprehensive review identifies PER as the dominant pattern for verifiable deployment in test generation, anomaly detection, code documentation, and more. Reviewer feedback is tied to automated, executable predicates, enabling closed-loop convergence on verifiable outcomes (Apostolou et al., 14 May 2026).
• Autonomous Research: ARIS applies PER with adversarial multi-agent reviewer routing, integrating skill libraries, persistent research artifacts, and a multi-stage audit cascade encompassing both experimental integrity and manuscript verification (Yang et al., 4 May 2026).

5. Traceability, Accountability, and Diagnostic Protocols

Role specialization in the PER pipeline enables fine-grained monitoring:

• Trace Logging: Each PER stage logs its input, output, and artifacts (e.g., prompt, intermediate answers, error flags, final answer), with raw trace records supporting a posteriori error attribution (Barrak, 8 Oct 2025). The schema identifies error origin (planner, executor, or critic), supporting systematic debugging.
• Structured Handoffs: Accountable PER pipelines enforce strict minimal handoffs (prompt plus current answer/artifact), avoiding untracked state or leaky context between roles.
• Blame Assignment: The point of initial error that propagates to the final answer is algorithmically assigned by traversing logs, enabling targeted model swaps or workflow refinements.
• Repair/Harm Metrics: For each stage $\Pi = P(G; C, f)$7, repair and harm rates are formally defined:

$\Pi = P(G; C, f)$8

allowing quantitative assessment of each agent’s efficacy and risk (Barrak, 8 Oct 2025).

6. Limitations, Industrialization, and Design Guardrails

The PER pattern, while dominant, faces contextual limitations primarily in non-determinism of LLM outputs, context window saturation, hallucination of tools/actions, and unbounded action spaces. Industrial best practices for mitigation include:

• Action Space Bounding: Permit only approved actions/tools in executor outputs, enforced via schema or static parsing (Apostolou et al., 14 May 2026).
• Structured Output Grammar: Require agent outputs to conform to formal grammars (e.g., JSON schemas with validation).
• Executable Feedback Loops: Reviewer outputs are grounded in runnable test harnesses, automated metric thresholds, or deterministic audit passes, preventing spurious task completions.
• Partitioned Context: Use of context-saving, preview mechanisms, and external persistent storage to manage context budget in executor–reviewer exchanges (Molinari et al., 3 Dec 2025).
• Adversarial Reviewer Pairing: Pairing executor and reviewer agents from different model families (e.g., Claude executor, GPT reviewer) provides diversity, increasing the probability of catching errors and unsupported claims (Yang et al., 4 May 2026).
• Verifiability-anchored Convergence: Only verified, test-passing artifacts exit the pipeline. This bounds system autonomy and reproducibility, accommodating enterprise and industrial reliability requirements (Apostolou et al., 14 May 2026).

A plausible implication is that, as PER pipelines mature, further industrialization will likely entail more granular, role-specific contracts; expanded reviewer autonomy; and hierarchical compositions, with each layer’s outputs entering downstream, context-pruned PER cycles.

7. Impact, Comparative Results, and Research Directions

Empirical studies demonstrate that PER architectures outperform monolithic or unreviewed agent systems on measurable metrics of reliability, robustness, and interpretability:

• Accountable handoffs increase worst-case pipeline accuracy by up to +36 percentage points over unmonitored chains (Barrak, 8 Oct 2025).
• Each reviewer module (syntax/semantic repair, ensemble verification, rubric-based scoring) adds incremental robustness and reduces error propagation in complex, long-horizon tasks (Xu et al., 9 Jun 2025, Han et al., 5 May 2026).
• Industrial adoption is highest in domains where reviewer-verifiable, executable outputs are possible, such as test generation, deployment, codebase analysis, and anomaly detection. Earlier-stage, open-ended tasks lacking such ground truth remain challenging for autonomous deployment (Apostolou et al., 14 May 2026).

Current trends prioritize dynamic role casting, structured trace auditing, and escalation/fallback from stage to stage (e.g., local vs. global plan repair, direct verification calls), with research directions targeting further increases in generalization, adversarial robustness, and seamless integration with automated orchestration layers.

---

Essential citations: (Molinari et al., 3 Dec 2025, Apostolou et al., 14 May 2026, Barrak, 8 Oct 2025, Xu et al., 9 Jun 2025, Han et al., 5 May 2026, Yang et al., 4 May 2026).