Skip to content

Knowledge base

The thin-harness ecosystem

Why Claude Code keeps its core loop small, and what that means for anyone building plugins, MCP servers or other extensions on top of it.

Written by

Updated:

Tools 4

Our position. We read a harness as a list of things the model cannot yet do reliably on its own (opens in a new tab). When a model improves at one of them, the matching component can shrink or go: a harness that never shrinks is a harness nobody is reading.

The bet behind the design is that model improvements beat harness work. Claude Code’s creator sums it up as “never bet against the model” (Plan mode and CLAUDE.md interview (opens in a new tab)), and the ecosystem grows in the space the core leaves open.

What does a thin harness mean for extensions?

  • The core harness delegates planning and adaptation to the model instead of encoding a large workflow engine.
  • Plugins, MCP, SDKs, and language servers carry specialization outside that core loop.
  • Common third-party patterns may become native features, so extension choices need regular review.

Why does Claude Code keep its harness thin?

Claude Code itself contains surprisingly little product logic. Its architecture is a single while loop over tool calls, no DAGs, no intent classifiers, no RAG pipeline; the design philosophy is “less scaffolding, more model,” so improvements in the model flow straight through to the tool (Claude Code). The team behind it confirms this: Claude Code is a lightweight shell where the model does almost all the work, the team writes as little business logic as possible, and around 90% of Claude Code is written with Claude Code itself (The Pragmatic Engineer - How Claude Code is built (opens in a new tab)). The harness shrinks as the model improves: with the Claude 4 models the team deleted roughly half the system prompt, and every release lets them remove more scaffolding, tools, and prompting (The Pragmatic Engineer - How Claude Code is built (opens in a new tab)). The trend has since gone further than halving: as of July 2026 the team reports having deleted 80% of the Claude Code system prompt, again on the reasoning that the model no longer needs the instruction, and that each model generation also invalidates the harness’s assumptions about output format, with engineers on the team now preferring HTML over Markdown because the model turned out markedly better at it than expected (The Pragmatic Engineer - How software gets built at Anthropic (opens in a new tab), July 2026, paid).

The same harness ships as the Claude Agent SDK, formerly the Claude Code SDK, and practitioners building on it reduce every agent to four levers, context, model, prompt, and tools; overwriting the SDK’s system prompt discards Claude Code’s built-in behavior entirely while appending only extends it (Building agents on the Claude Code SDK (opens in a new tab), 22 September 2025).

A March 2026 source leak put numbers behind the thin-harness story. Roughly 500k lines of orchestration logic shipped through source maps, no model weights, showing that fewer than 20 of the 60+ built-in tools are enabled by default and that subagents run a fork-join model over the KV cache, inheriting the parent’s full context without recomputation, which makes parallelism nearly free (AINews: The Claude Code Source Leak (opens in a new tab), March 2026). Anthropic answered with DMCA takedowns while a fork reached 32.6k stars and attackers registered malicious npm packages (color-diff-napi, modifiers-napi) targeting people compiling the leaked tree (AINews: The Claude Code Source Leak (opens in a new tab), March 2026).

The design stance is explicit in the interviews: Claude Code’s creator keeps a framed copy of the bitter lesson on the wall, summarizes it as “never bet against the model” (Plan mode and CLAUDE.md interview (opens in a new tab), 1 April 2026), and credits the step changes in Claude Code’s capability to model releases rather than harness work, “it was just the model” (The origin story of Claude Code (opens in a new tab), 2 June 2026).

A thin harness pushes customization outward, and Anthropic formalized that with plugins: bundles of slash commands, agents, MCP servers, and hooks that install with a single command, toggle on and off to keep system prompt context lean, and distribute through marketplaces that require nothing more than a git repository with a .claude-plugin/marketplace.json file (Anthropic - Customize Claude Code with plugins (opens in a new tab)). This low barrier is why the plugin ecosystem contains everything from one-person skill collections to official vendor plugins from OpenAI, Vercel, Hugging Face, and Trail of Bits. At the far end of the spectrum sit whole configuration frameworks that wrap the harness into what they call a structured development platform through behavioral instruction injection: SuperClaude bundles 30 slash commands, 20 agents, 7 behavioral modes, and 8 MCP servers, while Everything Claude Code ships 60 agents and 230 skills alongside hooks, MCP configs, and a session state store (SuperClaude (opens in a new tab); Everything Claude Code (opens in a new tab)).

By March 2026 the harness layer itself was contestable: the Pi agent harness lets practitioners define directory structures beyond .claude, parse custom frontmatter (expertise, skills, model) into the system prompt at boot, and override the system prompt entirely for non-coding agents; one showcase runs an adversarial board of revenue, contrarian, and moonshot agents in parallel under an Opus CEO, each maintaining accumulated per-agent “expertise” files as problem-specific memory and generating SVG arguments to persuade the orchestrator, and notes that Anthropic’s flat 1M-context pricing with no long-context premium is what makes such teams affordable, with quality degrading from around the 250k mark (A CEO-and-board multi-agent team on the 1M context window (opens in a new tab), 23 March 2026).

One harness pushes contestability a step past a system-prompt override. Its plugins can change the mechanics of the agent loop itself, such as how parallel tool calls are batched, rather than only reconfiguring the prompt around a fixed loop, and it can author and hot-reload such a plugin from a description given inside the running session. The security model has not followed: every installed plugin receives full shell and filesystem access by default with no per-plugin scoping, which the reviewer names as the price of that openness (Why DeepSeek Harness Just Became The Fastest Growing Github Repo EVER (opens in a new tab), 19 August 2026, practitioner).

MCP amplifies the effect beyond Claude Code. Because the protocol standardizes how any client connects to tools and data, every MCP server built for one assistant becomes a free plugin for anything that speaks MCP, an accidentally universal plugin ecosystem that nobody centrally planned (Works on My Machine - MCP: An (Accidentally) Universal Plugin System (opens in a new tab)). The same dynamic shows up in the cross-tool positioning of most entries in the Claude Code ecosystem catalog: skills and rules libraries increasingly target Claude Code, Cursor, Copilot, and Codex simultaneously.

The ecosystem is not only third-party. Anthropic ships its own free training through Anthropic Academy, including Claude Code 101, Claude Code in Action, and dedicated courses on agent skills and subagents (Anthropic Academy (opens in a new tab)). Its AI Fluency course frames effective collaboration as four practices, Delegation, Description, Discernment, and Diligence, the 4D framework, a vendor-authored baseline for the human half of the loop (Anthropic - AI Fluency: Framework & Foundations (opens in a new tab)). Claude Code also increasingly ships native equivalents of popular third-party categories: it builds auto memory as it works, saving learnings like build commands and debugging insights across sessions with no configuration, overlapping what third-party memory plugins provide (Anthropic - Claude Code overview (opens in a new tab)).

By December 2025 Language Server Protocol support was baked in, installed per language via /plugin; with an LSP attached Claude resolves exact references and function signatures instead of falling back to grep, absorbing what practitioners previously wired up through the Serena MCP server (Claude Code LSP support walkthrough (opens in a new tab), 22 December 2025; A Guide to Claude Code 2.0 (opens in a new tab), December 2025). Memory has reportedly grown into a four-layer system, CLAUDE.md (user-written), auto memory, background session memory, and a periodic “auto dream” consolidation that fires only once 24 hours and five sessions have both passed, runs read-only against project code, greps local transcripts rather than reading them whole, and prunes MEMORY.md to the 200-line startup budget; one observed run consolidated 913 sessions in roughly 8-9 minutes (vendor-described, and corroborated by the March 2026 source-leak analysis, which found a MEMORY.md index, topic files loaded on demand, searchable transcripts, and an autoDream mode in the code; Claude Fast – Claude Code Dreams (opens in a new tab), March 2026; AINews: The Claude Code Source Leak (opens in a new tab), March 2026).

Practitioners now track this absorption as a running pattern, native Code Review displacing Greptile and CodeRabbit, native voice dictation displacing Wispr Flow, the Tasks system displacing Beads, plan mode displacing Agent OS, and the Agent SDK displacing LangChain and CrewAI, to the point that shipping a thin skill or plugin gets likened to a YC AI-wrapper startup (Claude Code Best Practice (opens in a new tab)).

Resources

Harness contestability