This issue covers May 11 to May 17, 2026 (Asia/Taipei). If you think of “agents” as a new software stack, the open‑source story this week is: protocols are getting platformized, capabilities are getting packaged, and reliability/security are moving left into tooling.

Five practical signals worth tracking:

1) MCP becomes platform security plumbing: secret scanning via GitHub MCP server

The most compounding engineering signal this week: GitHub is tying MCP server integration into secret scanning, making “agent checks before commit/PR” a first‑class platform path. That pushes security capabilities into a managed, remote service layer (rules maintained upstream; teams consume it through MCP).

• InfoQ: GitHub Expands Secret Scanning with General Availability of MCP Server Integration (2026-05-12)
• GitHub Docs: Scan for secrets with the GitHub MCP server
• GitHub repo: github/github-mcp-server

For builders: pre‑commit/pre‑PR secret scanning should be the baseline for “vibe coding,” otherwise you’re deferring incidents to CI or production.

2) Multi-agent orchestration gets packaged: a repo-as-marketplace model

wshobson/agents is a useful “packaging” signal: it organizes multi‑agent collaboration, workflow orchestrators, skills, and commands as installable, granular plugins—explicitly optimizing for composability and lower context/token overhead.

• Repo: wshobson/agents

The value is less about “more prompts” and more about standardizing boundaries: plugin structure, skill loading, workflow decomposition, and composable units that can become a reusable ecosystem.

3) Memory becomes infrastructure (and a governance problem): agentmemory

One of the biggest practical blockers for real agent usage is “it forgets what it did.” agentmemory is a representative approach: persistent memory with MCP/plugin integration and explicit host integrations.

• Repo: rohitg00/agentmemory

My read: memory layers will quickly shift from “UX improvement” to “governance”: retention, deletion, secret/PII redaction, audit logs. Memory without policy becomes a new security surface.

4) Browser/DevTools control is a distinct lane: chrome-devtools-mcp + browsercode

If agents are going to deliver work, browser automation and DevTools instrumentation are unavoidable. Two representative directions:

• DevTools as agent tools: ChromeDevTools/chrome-devtools-mcp
• Browser-native agent framework: browser-use/browsercode

The hard problem isn’t “clicking.” It’s state consistency and replayability: DOM/network traces, retries, session contamination, and debugging primitives that can survive long horizons.

5) (Background) cloud vendors productize skills/plugins/MCP: AWS Agent Toolkit

This was primarily announced in the prior week, but it continues to shape the conversation: AWS is bundling MCP servers, skills, and plugins as an “Agent Toolkit,” with official docs and repo structure for safer cloud operations through agents.

• GitHub: aws/agent-toolkit-for-aws
• Docs: AWS Agent Toolkit documentation

Bottom line

Open source is building the “platform + governance layer” for agents:

• MCP’s value is moving from “tool connectivity” to “platform capability access” (e.g., secret scanning).
• Orchestration is getting packaged into reusable plugin/skill ecosystems.
• Memory is becoming foundational—and governance becomes mandatory.
• Browser/DevTools control is a necessary lane, where replayability will decide reliability.

Watchlist for next week

• Will GitHub MCP expand beyond secret scanning into broader supply‑chain/policy workflows?
• Will multi‑agent orchestration projects converge on shared interfaces (skills/hooks/MCP metadata)?
• Will memory layers standardize governance primitives (retention, redaction, audit)?
• Will browser/DevTools control adopt a common replayable trace format?