This article covers 2026-06-29 through 2026-07-05 in Asia/Taipei time. The weekly read: AI agents are moving from “can it demo?” to “can it be deployed under control?” Industry is dealing with cost, safety, and field implementation; research is decomposing evals, tool architecture, and memory; open source is turning terminals, assistant gateways, skills, and professional toolkits into new control planes.

This week’s deep dives:

• Industry weekly: Claude Sonnet 5, Fable 5, Claude Science
• Research weekly: PACE, MCP Patterns, WorldEvolver
• Open-source weekly: Warp, OpenClaw, Agent Skills

1. Industry: launches now include price, safety, and governance

Claude Sonnet 5 presents an agentic model through effort levels and cost/performance curves, with explicit pricing and evaluation across agentic search, computer use, and coding. Fable 5 redeployment reads like a frontier-model incident-response case: export controls, suspended access, a new classifier, a 99% blocking claim for a reported bypass, government pre-release testing, and a cross-company jailbreak severity framework.

Claude Science shows what a vertical agent workbench can look like: tools, databases, HPC, artifact provenance, reviewer agents, and domain skills in one environment. NVIDIA’s agentic RL guide turns agent improvement into a loop of evals, environments, verifiers, GRPO, and held-out validation.

2. Research: agent capability is becoming measurable components

PACE uses low-cost proxy evals to estimate expensive agent benchmark performance, giving model routing and checkpoint triage an early signal. MCP Server Architecture Patterns gives the MCP server ecosystem a taxonomy of five architecture patterns and several anti-patterns, emphasizing that tool count, auth, versioning, and observability directly affect reliability. WorldEvolver puts world models and memory into long-horizon planning, letting agents adjust foresight from prediction-observation mismatches.

Together, these papers decompose agents into eval, tool boundary, and memory/world-model problems rather than treating the system as one opaque demo.

3. Open source: the control-plane race intensifies

Warp open-sourced an agentic development environment built from the terminal. OpenClaw represents a local-first, multi-channel personal assistant gateway. agent-skills and .NET skills show procedural memory becoming a versioned asset. MATLAB and Simulink are making professional engineering tools agent-ready.

At the same time, reports on indirect prompt injection, poisoned MCP tool descriptions, shell injection, and package verification remind us that a control plane without sandboxing, policy, audit logs, and confirmation flows just connects untrusted content to trusted tools.

Weekly Synthesis

AI agents are entering controlled deployment. Capability is only the first layer; the real competition is cost curves, evaluation methods, tool boundaries, memory updates, vertical workbenches, field delivery, and safety governance. The question for next week is whether these control layers become defaults rather than expert-only patches.

Watchlist

• Whether Claude Sonnet 5 remains attractive for agent workloads after introductory pricing ends.
• Whether the Fable 5 / Glasswing jailbreak framework becomes public and adopted by other labs.
• Whether PACE and MCP Patterns influence SDKs, registries, or CI eval pipelines.
• Whether Warp and OpenClaw make security and permissions defaults rather than documentation.
• Whether official skill libraries become a new distribution format for coding-agent ecosystems.