--- title: "Eleven Coding Agents, One Architecture: What a Source-Code Study Means for Enterprise Teams" slug: "coding-agent-harness-source-code-study-eleven-agents" author: "Jaione Amigot" date: "2026-11-06 09:00:00" category: "Premium" topics: "coding agents, agent harness, harness engineering, claude code, openai codex, gemini cli, agent skills, skill.md, mcp, agent permissions, sandbox, enterprise ai architecture, source code ownership" summary: "A Wavestone AI Lab study of eleven coding agents' source code found they converge on seven subsystems, use no vector search for code, and differ most in safety and permissions, which is where enterprise teams should spend their evaluation." banner: "" thumbnail: "" linkedin: | Four researchers at Wavestone AI Lab read the source code of eleven coding agents, about four million lines of Python, TypeScript and Rust, including Claude Code, Codex CLI, Gemini CLI, OpenHands and Aider. They found seven subsystems, built independently, that keep reappearing: the agent loop, model integration, tools, memory and context, safety and permissions, orchestration and extensibility. Two absences held across the corpus. No agent runtime imports a general-purpose agent framework, and none retrieves code with vector embeddings. Search is deterministic: ripgrep, tree-sitter, glob and Markdown context files. Skills, the SKILL.md format, ship in 9 of the 11. MCP ships in 8. The part to read twice is safety. The approaches go from four-layer stacks with native sandboxing to agents that ship no sandbox at all. So the loop is not where agents differ. What separates them is what the agent may touch, who sets that policy, and whether your skills can move with you when you switch. #iblai #CodingAgents #AIAgents #AgentSkills #EnterpriseAI --- ## The Short Answer **A source-code study of eleven coding agents found they share seven subsystems and the same retrieval approach, so the agent loop is no longer the differentiator. Enterprise teams should evaluate permissions, sandboxing and portable skills instead. With ibl.ai you own all the code and the data, and our coding-agent skills are plain Markdown.** The study is [Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents](https://arxiv.org/abs/2609.00006), an 83-page source-code study of eleven systems submitted to arXiv on 15 July 2026. Its authors are Paul Barbaste, Tristan Darrigol, Germain Vu and Tom Wiltberger of Wavestone AI Lab. Barbaste is also with Inclusive Brains. ## What did the Wavestone AI Lab study of coding agents actually examine? The Wavestone AI Lab team read the source of eleven production coding harnesses, roughly four million lines of Python, TypeScript and Rust. A twelfth system, a meta-harness called Omnigent, served as a contrast point. The eleven are Claude Code, Codex CLI, Gemini CLI, Mistral Vibe, OpenHands, Aider, Mini-SWE-Agent, Hermes, Pi, OpenCode and OpenClaw. The paper reports 13 cross-cutting observations, a catalogue of 29 design patterns and 18 design recommendations. Its working definition is the useful part for buyers: an agent is a model plus a harness, the runtime that connects the model to tools, context, safety controls, orchestration and extensions. ## Do all coding agents use the same architecture? The eleven coding agents converge on seven canonical subsystems, which the paper's first table names as follows. Convergence is not sameness, though: the same subsystem is built very differently from one agent to the next.
SubsystemWhat it does
Agent loopRuns the cycle of prompt, model response, tool call and result
LLM integrationConnects to one or more model providers
Tools and actionsReads and edits files, runs commands, calls services
Memory and contextDecides what stays in the context window and what is compacted
Safety and permissionsDecides what the agent may do without asking
OrchestrationCoordinates subagents and multi-step work
ExtensibilitySkills, hooks, plugins and MCP servers
The authors also observe that convergence is turning into imitation. Codex adopts Claude Code's hook vocabulary verbatim and ships an importer for its sessions and settings. ## Do coding agents use vector databases to search code? None of the eleven coding agents retrieves code with vector embeddings. The paper describes retrieval as deterministic: ripgrep, tree-sitter, glob, and Markdown context files the agent discovers automatically. Aider adds a ranked repository map. The second absence is just as plain. No agent runtime in the corpus imports a general-purpose agentic framework. Each team wrote its own loop, and the two absences held when the authors tripled the size of the corpus. For a team weighing a code-search index or an agent framework as the foundation of its developer tooling, the production evidence points the other way. ## Where do coding agents actually differ? Coding agents differ most in safety and permissions. The paper finds Claude Code uses three permission layers, Codex a four-layer stack with native sandboxing and an approval reviewer, and Gemini CLI four approval modes plus a sandbox. At the other end, Aider and Mini-SWE-Agent carry minimal safety, the paper's sandbox table lists none for Aider, Pi, OpenCode or OpenClaw, and OpenCode is permissive by default. Pi argues that a documented absence of sandboxing is itself a security position. Context handling varies as much. The paper counts seven systems that compact at a threshold, each in its own way, while Mini-SWE-Agent keeps an unbounded linear history. So two agents with the same seven boxes on a diagram can carry very different risk once they reach a real repository and a real terminal. ## What should enterprise teams evaluate in a coding agent instead? With the loop shared across eleven coding agents, evaluate what is not shared. Four questions do most of the work: 1. **What can the agent reach?** Test the default network, filesystem and credential exposure, not the configuration you hope people set. 2. **Who sets the policy?** Behavioural policy is moving from prompt text into configuration, the paper finds. Ask whether an administrator can set it centrally. 3. **Do your skills travel?** Skills ship in 9 of 11 agents and MCP in 8, so knowledge written as SKILL.md files survives a change of tool. 4. **Is there a record?** Ask for the log of what the agent ran, not only what it said. The paper's conclusion is that in the first half of 2026 the coding harness turned from a tool into a platform. Platforms are where lock-in lives, which makes the fourth question about ownership. ## How does ibl.ai fit the harness picture for coding agents? ibl.ai works with the harness your developers already run rather than replacing it. [Agentic Vibe](/blog/what-is-agentic-vibe-enterprise-ai-apps-coding-agent) is an MIT-licensed set of more than 150 skills that installs into Claude Code, Codex, Cursor, OpenCode or GitHub Copilot with one command. Three of those five, Claude Code, Codex and OpenCode, are in the Wavestone corpus. The skills are Markdown, so they follow the convergence the paper documents instead of a format only we read. Inside the platform, [Agent Skills](/docs/os/agent-settings/skills) are written playbooks an agent loads when relevant. The [Sandbox](/docs/os/agent-settings/sandbox) panel gives an agent a Linux virtual machine that starts with no network at all. An administrator opens it per agent to package registries, the public internet, or a custom allowlist of named hosts. The same panel can connect an agent to a registered OpenClaw instance, one of the eleven harnesses the study read. Since the paper lists no sandbox of OpenClaw's own, that is where the platform's controls earn their keep. [Transcripts](/docs/os/analytics/transcripts) and [Audit](/docs/os/analytics/audit) keep the record. ## Who should own the harness that runs your coding agents? If the harness is now a platform, owning it matters more than which model sits inside it. With ibl.ai you own all the code and the data, model-agnostic across any LLM, with no per-seat pricing. You can deploy anywhere: your own cloud, on-premise, GovCloud or a fully air-gapped network. Syracuse University runs ibl.ai this way, among 1.6M+ users across 400+ organizations. ibl.ai is family-owned and operated from New York, NY. Start from the [Agentic Vibe page](/product/agentic-vibe), or [book a pilot](/contact). *Source: Barbaste, Darrigol, Vu and Wiltberger, [Harness Engineering: Anatomy, Architecture, and Evolution of Coding Agents](https://arxiv.org/abs/2609.00006), arXiv:2609.00006, July 2026. Figures on subsystems, retrieval, adoption and permissions are the paper's own; we did not re-run its analysis. Product facts are from our documentation, linked above.* ## Why does owning the AI stack matter? **ibl.ai is the agentic AI platform where you own all the code and the data. You self-host the entire stack inside your own perimeter, run it model-agnostic across any LLM and switch anytime, and pay by usage with no per-seat pricing — so you can deploy anywhere: your cloud, on-premise, GovCloud, or fully air-gapped.** - **You own all the code and the data.** Full source code under a perpetual license, running on your infrastructure. Not API access to someone else's platform — the stack itself is yours. - **Model-agnostic.** Run any LLM — Claude, GPT, Gemini, Llama, Command, or your own fine-tune — and switch providers without rewriting the platform. - **No per-seat pricing.** Usage-based billing against a budget cap you set. Cost tracks what your organization actually uses, not how many people you employ. - **Deploy anywhere.** Your cloud, your VPC, on-premise, GovCloud, or a fully air-gapped network with no outbound connectivity. 1.6M+ users across 400+ organizations run the platform this way, including NVIDIA, MIT, and Syracuse University. ibl.ai is family-owned and operated from New York, NY — a U.S.-headquartered, domestically-owned long-term partner, not a vendor that sells licenses and moves on.