This package covers 2026-07-06 through 2026-07-12 in Asia/Taipei time (W28). All three tracks point to one conclusion: agent capability is becoming a platform feature, while long-horizon reliability, tool permissions, and observability constrain production value.

This week’s three deep dives

• Industry: Models Become Agent Runtimes (GPT-5.6, Muse Spark 1.1, GPT-Live)
• Research: Long Tasks Fail as Systems (Failure Taxonomy, OSWorld 2.0, PlanBench-XL)
• Open Source: Skills, MCP, and Harnesses Become Composable Infrastructure

One picture: capability descends, responsibility rises

GPT-5.6 puts persisted reasoning, programmatic tools, and multi-agent orchestration into the model platform. Muse Spark 1.1 combines long context, subagents, MCP, computer use, and coding harnesses. GPT-Live and Muse Image turn voice and media interfaces into actors. The standard agent runtime is expanding quickly.

Research shows why that does not guarantee outcomes. OSWorld 2.0’s professional workflows, PlanBench-XL’s failing tools, and a cross-benchmark failure taxonomy all show that planning, state, recovery, coordination, and measurement must be evaluated together. Completing a step is not the same as owning a result.

Open source supplies the other half: skills preserve procedures, MCP exposes tools, and harnesses manage configuration. Each capability also adds prompt-injection, credential, irreversible-action, and supply-chain risk. The next stack needs capability manifests, version pinning, traces, checkpoints, and human escalation.

The practitioner takeaway

Do not ask only which model is best. Measure end-to-end completion, cost per success, recoverability, intervention points, permission scope, and whether the trajectory can be reproduced. Models, benchmarks, and open-source infrastructure are converging on the same systems problem.

Watchlist

• Shared trace and state standards across native multi-agent APIs.
• Long-horizon benchmarks becoming enterprise regression suites.
• Provenance, signatures, and permission manifests for Skills and MCP.
• Confirmation and recovery design for sensitive voice/multimodal actions.
• Published completion, cost, and incident evidence—not demo counts.