What does software engineering become when no one writes the code?

An OpenAI team shipped close to a million lines in five months with zero hand-written, all by Codex. It isn't model magic but engineering that moved up a layer: from writing code to building the environment the agent runs in, with legible logs, a repo as the system of record, and linter-enforced architecture. They warn it took heavy investment and won't generalize for free.

https://benjaminhan.net/posts/20260626-harness-engineering/?utm_source=mastodon&utm_medium=social

#AgenticSystems #OpenAI #AI

What if you searched over an agent's code, not just its prompts? Automated Design of Agentic Systems does exactly that: a meta agent writes new agent scaffolds in Python and keeps the ones that score well on a task. The discovered designs beat hand-built baselines, and they keep working on new tasks and new base models that the search never optimized for.

https://benjaminhan.net/posts/20260625-adas/?utm_source=mastodon&utm_medium=social

#AgenticSystems #Metacognition #LLMs #AI

How much of a coding agent is the loop, and how much is the machinery around it? A source-level read of Claude Code's TypeScript finds the loop almost trivial: call model, run tools, repeat. Almost all the code is around it: a deny-first permission gate, a five-layer compaction pipeline, four graduated extension mechanisms, isolated subagents. The open-source OpenClaw gateway answers the same questions differently.

https://benjaminhan.net/posts/20260623-claude-code-design-space/?utm_source=mastodon&utm_medium=social

#Paper #AgenticSystems #Claude #Coding #AI

Where does an AI math agent get its ability, the model or the orchestration around it? In the first large-scale test of formal proof search on open problems, an agent closed 9 of 353 Erdős problems in Lean. In its own ablation, a plain generate-and-verify loop solved all nine, while smaller models and the specialized prover alone solved nothing.

https://benjaminhan.net/posts/20260606-ai-formal-proof-search/?utm_source=mastodon&utm_medium=social

#AI #AIforScience #Mathematics #AgenticSystems

Where should an agent run relative to its sandbox? LangChain shipped Deep Agents, a model-agnostic take on the harness pattern behind Claude Code. Its key difference from the Claude Agent SDK is exactly that question: Deep Agents can run outside the sandbox and drive it as a tool, so credentials stay off the sandbox and the security boundary wraps execution, not the agent. That decoupled shape is where production is heading.

https://benjaminhan.net/posts/20260529-langchain-deep-agents/?utm_source=mastodon&utm_medium=social

#AgenticSystems #AIEngineering #AI

If hundreds of subagents check each other and agree, is the answer right? Claude Code can now write an orchestration script that fans a task across them in parallel, with other agents refuting each finding until they agree. But convergence is not correctness. They share one base model and context, so they share blind spots, and wrong answers converge as cleanly as right ones. It only holds if the refuters are genuinely adversarial.

https://benjaminhan.net/posts/20260529-dynamic-workflows-claude-code/?utm_source=mastodon&utm_medium=social

#AgenticSystems #AIEngineering #AI

Should you pick a boring language like Go so coding agents write more reliable code? A widely-shared argument says low-variance ecosystems beat fragmented ones like JavaScript or Python. The likelier driver is corpus size, not variance: models are strongest on the most-represented languages, the fragmented ones. And betting on the median buys reliability where it was already cheapest, not on the unfamiliar work that costs time.

https://benjaminhan.net/posts/20260529-use-boring-languages-with-llms/?utm_source=mastodon&utm_medium=social

#AgenticSystems #SoftwareEngineering #AI

What are the effective ways to use Claude Code in large codebases? Anthropic's recommended order to build up the harness is CLAUDE.md, hooks, skills, plugins, LSP, then MCP. The most common failure they see is reaching for MCP first, before the basics work. For outcomes, the harness matters more than the model swap.

https://benjaminhan.net/posts/20260528-claude-code-large-codebases/?utm_source=mastodon&utm_medium=social

#AI #AgenticSystems #SoftwareEngineering #Anthropic

Singapore Researchers Harmonize Diverse SIEMs with Agentic Rule Translation

Imagine having multiple Security Information and Event Management platforms working in perfect harmony - Singapore researchers have made this a reality by developing a game-changing approach called agentic rule translation, enabling seamless interoperability between diverse SIEMs.…

https://osintsights.com/singapore-researchers-harmonize-diverse-siems-with-agentic-rule-translation?utm_source=mastodon&utm_medium=social

#SiemInteroperability #AgenticSystems #SecurityInformationAndEventManagement #Singapore #ResearchAndDevelopment