Building a C compiler with a team of parallel Claudes

Summary

Nicholas Carlini, a researcher on Anthropic’s Safeguards team, describes an experiment in which 16 parallel instances of Claude Opus 4.6 worked autonomously on a shared Rust repository to produce a 100,000-line C compiler. Over roughly two thousand model sessions and $20,000 in API usage, the resulting compiler reached a 99 percent pass rate on major test suites and successfully built a bootable Linux 6.9 kernel for x86, ARM, and RISC-V targets, as well as several large open-source packages.

The central contribution lies in the scaffolding required to sustain long-running agent teams without continuous human oversight. Each agent runs inside its own Docker container mounted against a bare upstream Git repository; when an agent finishes a task it pushes its changes and immediately pulls the next available item. A lightweight locking scheme prevents concurrent work on the same file or bug, while extensive READMEs and progress logs allow freshly spawned agents to orient themselves quickly.

Test design proved decisive. Early in the project, agents could be assigned independent failing tests, but once the suite reached high coverage the remaining work collapsed into a single interdependent task—compiling the Linux kernel. To restore parallelism, the harness used GCC as an oracle: most files were compiled with the reference compiler and only a random subset with the agent-built compiler, allowing different agents to isolate and repair distinct defects. Additional agents were later specialized for duplicate-code removal, performance tuning, and documentation.

The compiler itself remains limited. New features frequently regressed existing functionality, and the model was unable to produce a compact 16-bit x86 backend that satisfied Linux’s size constraints for real-mode boot code; that phase still delegates to GCC on x86. These shortcomings illustrate both the current ceiling of fully autonomous agent teams and the value of the harness techniques developed to reach it.

01:00 · March 24, 2026

Harness design for long-running application development

This article provides highly actionable architectural patterns for product teams and builders developing autonomous AI agents. It offers concrete solutions to common LLM limitations like context degradation and self-evaluation bias, which are critical for Dutch AI engineering teams building robust, long-running applications.

Relevance 85 · Audience 95

02:00 · August 20, 2026

How monday.com transformed its platform into an agent-first product where humans and agents collaborate

This case study is highly relevant for product teams and builders as it provides a strategic blueprint for transitioning from superficial AI features to a native, agent-first architecture. It offers actionable insights into integrating LLMs like Claude into core workflows, which is highly applicable for Dutch SaaS companies and AI practitioners looking to drive sustained user engagement.

Relevance 75 · Audience 90

02:00 · June 24, 2026

Building effective human-agent teams

Provides actionable workflows, role definitions, and verification practices for Product Teams and Builders integrating agentic AI into real team processes, directly supporting implementation of new Claude capabilities.

Relevance 78 · Audience 85

02:00 · April 8, 2026

Scaling Managed Agents: Decoupling the brain from the hands

Highly actionable for Product Teams and Builders implementing agent workflows with Claude, including code-level interface patterns, security mitigations, and performance gains like reduced TTFT. Directly addresses model updates, harness evolution, and production observability.

Relevance 80 · Audience 85

01:00 · November 26, 2025

Effective harnesses for long-running agents

Highly actionable for Product Teams and Builders with concrete implementation guidance, exact model references (Opus 4.5), and workflow strategies directly applicable to agent development and observability.

Relevance 82 · Audience 88

02:00 · October 16, 2025

Equipping agents for the real world with Agent Skills

Directly actionable for Product Teams and Builders: provides concrete implementation patterns, evaluation guidelines, and code patterns for building specialized agents. Addresses lifecycle, observability via progressive loading, and risks like malicious skills.

Relevance 78 · Audience 85

21:12 · August 2, 2026

How to Become a Graph Architect With Zero Experience (Full Course)

Provides hands-on guidance on agentic systems and graph orchestration frameworks such as LangGraph, directly applicable to AI engineers building production multi-agent applications in the Netherlands and EU.

Relevance 55 · Audience 50

02:00 · July 22, 2026

How Outtake built a cyber investigator on Claude

This article provides a practical use case for Product Teams and Builders on how to leverage Claude Code and the Agent SDK to build long-running, autonomous AI agents. It offers valuable architectural insights for Dutch AI practitioners developing cybersecurity solutions or complex agentic workflows.

Relevance 75 · Audience 85

02:00 · July 20, 2026

Working at the frontier: How Rakuten builds agents overnight with Claude Fable 5

This article provides product teams and builders with insights into deploying long-running, autonomous AI agents using Claude Fable 5. It highlights practical enterprise strategies for balancing model intelligence with cost and managing human-in-the-loop constraints, which are highly applicable to Dutch AI practitioners scaling agentic workflows.

Relevance 75 · Audience 85

02:00 · July 16, 2026

Working with Claude Fable 5 in Claude Cowork

This article is highly relevant for product teams and builders as it provides actionable insights on integrating Anthropic's latest agentic model into complex workflows. Dutch AI practitioners can use these updates to enhance productivity, automate multi-step processes, and understand the operational nuances of Claude Cowork.

Relevance 85 · Audience 95

06:00 · July 11, 2026

Context Graphs for Proactive Enterprise Agents

High technical depth, novel proactive architecture, and complete reproducible implementation make it directly actionable for Dutch AI researchers and advanced enterprise practitioners developing agent systems.

Relevance 78 · Audience 92

06:00 · July 9, 2026

The Harness Effect: How Orchestration Design Sets the Token Economics of Enterprise Agentic AI

Provides actionable, reproducible insights into orchestration design that Dutch AI teams and enterprises can directly apply to lower costs of agentic systems; presents novel empirical findings and formalization suitable for advanced researchers.

Relevance 68 · Audience 82