Detailed Analysis
A developer building a large frontend codebase migration has released "neal," an open-source CLI orchestrator that formalizes a pattern many practitioners have arrived at independently: using one LLM to write code and a different LLM to review it. Built initially around OpenAI's Codex GPT-5.4 and Anthropic's Claude via their respective SDKs, neal now also supports OpenRouter, giving it access to 44 compatible models. The tool emerged from practical friction encountered during a real migration—549 commits and over 3,000 files changed—where the author found that simply instructing an agent to "keep working unless blocked" broke down over long task horizons. That observation led to a structured architecture: work is decomposed into discrete chunks via a planner/reviewer loop, each chunk is executed by a coder whose context is reset between chunks to avoid "context rot," and a separate read-only reviewer agent with a persistent, longer-running context evaluates the work before moving forward. A final pass reconciles the completed implementation against the original plan.
The most notable technical claim is the value of cross-model adversarial review. When the author ran 100 SWE-bench Pro cases that Codex had failed to solve independently through neal, using Codex in both coder and reviewer roles solved 8 of them, while swapping in Claude as the reviewer solved 15—nearly double. This is a modest but meaningful data point in a growing body of evidence suggesting that model heterogeneity in multi-agent systems produces better outcomes than homogeneous agent pairs, likely because different model families have different blind spots and failure modes that a same-model reviewer is less likely to catch. The author is candid that the benchmark improvement wasn't as dramatic as hoped, but frames it as validating the core thesis of the tool rather than the tool being a definitive breakthrough.
This project reflects a broader shift happening across the AI coding tools ecosystem in 2025-2026: orchestration and process engineering are becoming as important as raw model capability. As models like Claude and Codex/GPT variants get better at code generation, the bottleneck for autonomous coding shifts to context management, task decomposition, verification, and multi-agent coordination—problems that aren't solved by scaling model size alone but require systems-level thinking. Context rot, in particular, has become a well-recognized failure mode as agents attempt longer-running autonomous tasks, and resetting coder context while preserving reviewer context is a pragmatic mitigation that other agentic frameworks have also begun adopting in various forms.
The tool's design also underscores the increasingly commoditized, interoperable nature of frontier coding models. By supporting Claude Code, Codex CLI, and OpenRouter simultaneously, neal treats specific models as swappable components within a pipeline rather than as fixed dependencies—users can mix and match planner, coder, and reviewer roles across vendors based on empirical performance rather than lock-in. The author's own production setup, using Claude as planner/coder and Codex as reviewer, exemplifies this flexibility. Notably, the author openly acknowledges that neal's usefulness may be short-lived given how quickly model capabilities are advancing—an implicit admission that orchestration tools built to compensate for current model limitations (context degradation, self-review blindness, poor long-horizon planning) may become less necessary as underlying models improve. This tension, between building scaffolding for today's models versus anticipating tomorrow's more capable ones, is a recurring theme across the current wave of agentic coding tooling.
Read original article →