Eleven coding agents opened at the source: no runtime imports a general agentic framework, and none retrieves code with embeddings
Synopsis
A source-code anatomy of eleven production coding-agent runtimes separates the agent into the runtime layer around the model and compares implementations across seven subsystems. Across roughly four million lines of Python, TypeScript and Rust, no runtime imports a general-purpose agentic framework such as LangChain, LangGraph or AutoGen, and none retrieves code with vector embeddings, relying instead on hand-rolled async loops and deterministic retrieval through ripgrep and tree-sitter. The study also diffs the same systems across one quarter and observes behavioral policy migrating from prompt prose into configuration.
Interpretation
What separates coding agents is less the model than the runtime around it; that runtime decomposes into seven fixed subsystems, and every system from a roughly hundred-line research baseline to a million-line product CLI must take a position on each. Earlier benchmark surveys reported SWE-Bench scores without touching architecture, conceptual taxonomies described patterns without implementations, and the one concurrent source-level taxonomy excluded the provider-native systems that define the field's frontier. The eleven systems span three programming languages and three orders of magnitude in code size, four of them flagship products of major commercial model providers, with one meta-runtime analyzed separately as a contrast point.
Most of the code in production systems addresses safety, user experience, extensibility and increasingly clients and transport rather than task completion. The minimal system, Mini-SWE-Agent, implements all seven subsystems in roughly one hundred lines, and its self-reported SWE-Bench Verified results fall in the same range as systems three orders of magnitude larger. Roughly three-fifths of OpenCode's non-test source is its TUI, web, desktop and SDK clients rather than the runtime proper, and Codex's July tree devotes six-figure line counts to app-server transports, plugins and a realtime voice layer.
The extensibility standards have resolved: SKILL.md skills lead MCP in adoption, and ACP has acquired a third role, harness hosting. MCP had been treated as the standard interface connecting models to external tools, while the roles of skills and ACP had not been compared at source level. Nine of the eleven systems adopt SKILL.md skills and eight adopt MCP, ACP ships in six systems, and OpenHands can run Claude Code, Codex or Gemini CLI as interchangeable backends.
Perspective
The analysis covers the source of eleven coding runtimes pinned to July 2026 releases, plus Databricks' Omnigent as a meta-harness contrast point; Claude Code is read from a March 2026 source snapshot, and its July binary may differ. The study describes and compares how systems are built without benchmarking or ranking, so the conclusions serve readers who want to understand architecture and design trade-offs rather than which system performs better on tasks.
The study states that code-size figures are approximate line counts on pinned snapshots, not comparable across languages, and serve only to position systems on the simplicity-versus-capability axis; earlier SWE-Bench Verified scores were removed from the comparison table because they are self-reported, come from different model generations and configurations, and several predate current defaults. The longitudinal diff covers only the original eight systems, and three of the April edition's observations required substantive revision in place.
