The Rise of Multi-Agent Development

The most striking trend emerging from today's posts is the normalization of running multiple AI coding agents in parallel. What started as an experimental workflow has become standard practice for power users.

"gm to all multi clauders" — @cto_junior

"how many claude codes do you run at once?" — @pleometric

Jeffrey Emanuel (@doodlestein) shared his comprehensive "Flywheel" system, describing the later stages of development as "basically mindless machine tending of your swarm of 5-15 agents." His workflow emphasizes front-loading human effort into planning while relegating implementation to parallel agent execution.

@idosal1 announced building AgentCraft v1, managing "up to 9 Claude Code agents with the RTS interface." The gaming metaphor is apt—developers are increasingly treating AI agents like units in a real-time strategy game, coordinating multiple autonomous workers toward a common goal.

The Skills Ecosystem Expands

Claude Code's skills system is gaining traction across the ecosystem:

• Vercel released react-best-practices, installable via npx add-skill vercel-labs/agent-skills, containing performance rules and evals to catch regressions
• Trail of Bits published 17 security skills that @koylanai praised as "the beginning of something massive"
• @steipete's clawdbot continues to impress, with @LLMJunky describing tasks kicked off "from the Denny's parking lot" that coordinated repo research, documentation pulling, and migration planning

@koylanai made a bold prediction: "Every company with technical docs will ship Skill packages, not because it's nice to have, but because agents won't adopt your product without them. Agents (or humans) won't read docs; they execute Skills."

Platform Wars Heat Up

OpenAI: Open Responses Spec

OpenAI Developers announced Open Responses, an open-source specification for building multi-provider, interoperable LLM interfaces. Key features:

• Multi-provider by default
• Designed for real-world workflows
• Extensible without fragmentation

The pitch: "Build agentic systems without rewriting your stack for every model."

GitHub: Copilot Gets Memory

GitHub Copilot now has agentic memory in public preview:

• Learns repo details to boost agent, code review, and CLI help
• Memories scoped to repos, expire in 28 days
• Shared across Copilot features

Cursor: Better Bug Detection

"Cursor now catches 2.5x as many real bugs per PR." — @cursor_ai

The Planning Paradox

Jeffrey Emanuel's thread contained a critical insight that deserves attention:

"The one thing people seem to get wrong is ignoring what I say about planning or transforming their plan into beads. They make a slipshod plan all at once with Claude Code. Or they try to one-shot turning the plan into beads... Well, of course the project is going to suck and be a buggy mess if you do that."

The paradox: as AI makes coding faster, the value of human planning increases. Emanuel recommends spending "most of your energy and human time/focus on the markdown plan" and doing "at least 3 rounds of polishing, improving, and expanding" before letting agents execute.

AI Engineering's Runtime Problem

@ashpreetbedi highlighted a structural issue in the AI engineering stack:

"Claude Code shipped two years after function calling. Models have outpaced the application layer. We have frameworks to build agents, we have observability to trace them, we have evals to test them."

The article argues that the runtime layer—where agents actually execute—remains underdeveloped compared to model capabilities.

Generative Interfaces

Guillermo Rauch (@rauchg) shared a glimpse of "a world of fully generative interfaces" with the flow: AI → JSON → UI. This points toward a future where interfaces are dynamically generated rather than pre-built, raising questions about design systems, accessibility, and the role of frontend development.

Engineering Hiring in the AI Era

Mitchell Hashimoto offered a provocative take on interviewing:

"I think a really effective engineering interview would be to explicitly ask someone to use AI to solve a task, and see how they navigate. Ignore results, the way AI is driven is maybe the most effective tool at exposing idiots I've ever seen."

This suggests that AI proficiency isn't just a nice-to-have—it's becoming a core competency that reveals underlying engineering judgment.

The Humor Corner

@askOkara captured the zeitgeist:

"I saw a guy coding today. No Okara. No Cursor. No OpenCode. No Claude Code. He just sat there, typing code manually. Like a psychopath."

And Kent C. Dodds vindicated his earlier position on MCP:

"When everyone was saying MCP is doomed because context bloat, I was saying all you need is search. Feels good to have my bets validated once again."

Looking Ahead

Harrison Chase (@hwchase17) clarified LangChain's approach to agent memory: "We don't use an actual filesystem. We use Postgres but have a wrapper on top of it to expose it to the LLM as a filesystem." This abstraction pattern—making complex systems appear as familiar interfaces to LLMs—may become a key design principle.

@BlasMoros shared a prescient quote about software economics:

"LLMs have proven themselves to be remarkably efficient at [translation between human and computer language] and will drive the cost of creating software to zero. What happens when software no longer has to make money? We will experience a Cambrian explosion of software, the same way we did with content."

The multi-agent future isn't coming—it's here. The question now is how quickly the rest of the ecosystem catches up.