This article covers 2026-06-15 to 2026-06-21 in Asia/Taipei time. The open-source theme was clear: agent projects are expanding beyond chat UIs and simple tool calling into agent harnesses, skill registries, desktop/browser/mobile control, cloud skill packs, and embedded agents.

The GitHub numbers below are API snapshots checked on 2026-06-22. They are used as momentum and risk signals, not as exact week-over-week growth.

1. Hermes Agent and OpenClaw: personal agent shells drew huge attention

Two leads dominated the week:

• NousResearch/hermes-agent: about 199k stars, 35k forks, MIT license.
• openclaw/openclaw: about 380k stars, 79k forks; license metadata was not standardized in the API snapshot.

Both point toward a “personal agent shell” model: an agent as an entry point into OS, desktop, browser, and tools. The risk is also clear. When attention spikes this fast, issue triage, plugin safety, permission boundaries, and release discipline become product-critical.

2. Google ADK remained the low-friction enterprise path

google/adk-python continued to appear in this week’s leads. The API snapshot showed about 20k stars, 3.6k forks, Apache-2.0 licensing, and updates near the end of the window.

• GitHub: google/adk-python

ADK matters because official, code-first, evaluable, deployable tooling is what enterprises usually prefer when agents move toward production.

3. Browser and device harnesses moved closer to real operation

Two repos were especially relevant:

• browser-use/browser-harness: about 15k stars, 1.4k forks, MIT license.
• callstack/agent-device: a CLI for controlling iOS and Android devices, about 2.9k stars, MIT license.

This is the open-source answer to a research and product problem: browser, phone, and desktop operation are not the same as clean API or CLI tasks. Reliable real-interface control is becoming a core layer for useful agents.

4. Skill registries became a maintainability layer

VoltAgent/awesome-agent-skills and microsoft/azure-skills point toward skills and MCP/plugin configuration as maintainability infrastructure.

• GitHub: VoltAgent/awesome-agent-skills: about 26k stars, MIT license.
• GitHub: microsoft/azure-skills: about 1.2k stars, MIT license.

The pattern matches this week’s research track: skills should not live as loose prompts. They need versioning, review, retrieval, composition, and revocation.

5. Edge and vertical agents expanded the surface area

Three vertical projects were worth watching:

• espressif/esp-claw: IoT/embedded agent framework, about 1.6k stars, Apache-2.0 license.
• matlab/matlab-agentic-toolkit: agentic interface for scientific and engineering workflows.
• skalesapp/skales: personal desktop agent, about 1.1k stars; license metadata was not standardized in the API snapshot.

The implication: agent open source is moving into specific execution environments. That is useful, but it widens the security model to OS permissions, device bridges, credential storage, workflow replay, and plugin provenance.

Weekly View

The best open-source signal this week was not one repo. It was the pattern: agent harness + skills + device control + cloud skill packs are becoming reusable layers. Models will keep changing; durable value is moving into execution shells, skill lifecycle, and safety boundaries.

Watchlist

• Whether Hermes Agent and OpenClaw convert attention into stable releases and clear security models.
• Whether Google ADK adds stronger evaluation and deployment templates.
• Whether browser-harness and agent-device are adopted by higher-level coding and research agents.
• Whether awesome-agent-skills evolves into searchable schema rather than a curated list.
• Whether Azure skills turns MCP/skill packs into an enterprise cloud adoption pattern.