The Empire, Long Divided, Must Unite: Architectural Convergence in Three LLM Agent Harnesses

As autonomous coding agents become more capable, the focus of research has shifted from the underlying language models to the 'harness'—the surrounding code that manages context, tools, and state. This study investigates the architectural evolution of three prominent open-source agent harnesses: LangChain's deepagents (batteries-included), Earendil's pi (minimalist), and DeepSeek's dsh (plugin-based). By analyzing these systems at pinned commits and tracing their development history, the author identifies how these projects have moved from disparate starting points toward a common architectural structure.

The study finds that despite their different philosophies, all three harnesses have adopted a five-element 'middle form':

• Commoditized Loop: The execution loop is no longer a competitive differentiator; it is either kept small and readable or delegated to existing frameworks.
• Append-only Session Record: Systems have moved away from destructive history mutation in favor of replayable, append-only logs.
• Model Quirks as Data: Rather than hard-coding model-specific behaviors, harnesses now treat these as externalized data configurations.
• Progressive Disclosure: Context is managed by exposing only essential information to the model, with full details fetched on demand.
• Explicit Extension Seams: Architectures are increasingly modular, using clear interfaces to allow for plugin-based extensions rather than monolithic designs.

While the harnesses have converged on the elements necessary for functional, long-horizon operation, they share a significant blind spot: external verifiability. None of the studied harnesses provide a tamper-evident record that an independent third party could audit without trusting the runtime itself. The author argues that this is not an oversight but a 'predictive gap'—the next frontier for agent harnesses, particularly in domains where provenance, cost-tracking, and security are critical.