Twitter/X AI + Agentic Age: Eyes, Memory, and Logs.

Dates: March 29-April 11, 2026

Source note: I filtered my X bookmarks by tweet postedAt date. The bookmarks API gives the tweet date, not the exact date I bookmarked it, which is an annoying but useful distinction.

The end of March began with computer use in Claude Code and then immediately swerved into the Claude Code source leak, which gave the whole period a faint smell of melted plastic. The leak was less interesting as gossip than as x-ray: people were suddenly reading the machinery behind one of the most important developer tools in the world, and the machinery looked like what you would expect from a fast-moving product that had become infrastructure before it had time to become boring.

At the same time, the official direction was clear. Claude Code got computer use, Monitor, dynamic loops, /ultraplan, Word integration, and managed agents. Codex expanded memory with Chronicle. Cloudflare exposed Chrome DevTools Protocol through Browser Rendering. Google shipped Gemma 4. Karpathy wrote about LLM knowledge bases. The pattern was not “chatbots get features.” It was “agents get perception, memory, and runtime hooks,” which is a much stranger and more useful sentence.

The monitor tools were especially revealing. Polling is what you do when your agent is pretending to be patient. Monitoring is what you do when you admit the system has events, logs, failures, and waiting periods. Once agents can watch the world and wake up when needed, the loop stops being a chat loop and starts becoming a control loop.

Key movements

Computer use moved from demo to workflow. Claude Code could open apps, click UIs, and test what it built; Cloudflare and browser-tool builders made remote browser control cheaper and easier. The important shift is not that the agent has a mouse. It is that the agent can collect evidence from the same surface where the user would have discovered the bug.

Memory became a product surface. Codex Chronicle and Claude memory explainers made persistent context feel less like a prompt hack and more like infrastructure. Memory is a seductive word, but the actual product question is much less poetic: what gets stored, who can edit it, how it is scoped, and what happens when the agent remembers something false with great confidence.

Managed agents appeared as platform offerings. Claude Managed Agents and Vercel examples showed the labs moving toward hosted harnesses, not just models. This is the same move cloud made years ago: the value moves from the raw capability into the managed runtime around it.

Observability entered the agent loop. Monitor, browser traces, CDP access, and dynamic loops suggested that the next generation of agents would spend less time asking “are we done yet?” and more time listening for evidence. The interface becomes quieter when the agent can wait on a signal instead of repeatedly poking the user.

The leak reminded everyone that agent tools are software too. The Claude Code source-map incident was a mundane packaging failure with an unusually theatrical blast radius, which is often how real incidents work. Infrastructure rarely fails in mythic ways. It fails because some ordinary artifact crossed a boundary it was not supposed to cross.

Representative signals

Date	Actor	Why it mattered	Link
Mar 30	Claude	Computer use in Claude Code closed more of the build-test-observe loop.	View
Mar 31	Chaofan Shou	Claude Code source-map leak turned the agent stack into public archaeology.	View
Apr 2	Karpathy	LLM knowledge bases reframed token throughput as research infrastructure.	View
Apr 2	Cursor / Erik	Cursor design mode brought direct visual manipulation into the coding-agent surface.	View
Apr 8	Claude	Claude Managed Agents made agent platform a named product category.	View
Apr 9	Noah Zweben	Monitor let Claude create background scripts that wake the agent when needed.	View
Apr 10	Cloudflare Developers	Browser Rendering exposed CDP for remote browser control by agents.	View
Apr 10	Thariq	/ultraplan moved planning into an editable web artifact.	View

Engagement ledger

Tweet	Engagement
Claude computer use in Claude Code	59.5K likes, 16.1M views
Karpathy on LLM knowledge bases	58.2K likes, 20.8M views
Claude Managed Agents	57.1K likes, 21.5M views
Claude source leak	48.8K likes, 35.5M views
Claude advisor/executor strategy	38.5K likes, 4.7M views
Claude for Word	32.0K likes, 12.4M views
Theo on Claude Mythos	12.3K likes, 3.3M views
Boris Cherny NO_FLICKER mode	10.3K likes, 2.9M views

What changed in this increment was not just capability but posture. The agent was no longer simply answering, coding, or waiting. It was watching, remembering, inspecting, and waking itself when the world changed. That is a more serious kind of tool, which means it also needs a more serious kind of containment.