Agent Control Stack

Key Conclusions

• The common thread is not “better models.” It is that control surfaces are becoming first-class engineering objects.
• Agent quality now depends on harness, policy, and eval design.

Selected Signals

1. Evaluating alignment of behavioral dispositions in LLMs

• Source: Google Research Blog
• Date: 2026-04-03

Google Research turns behavioral alignment from a concept into reproducible, testable evaluation engineering.

2. Copilot organization custom instructions are generally available

• Source: GitHub Changelog
• Date: 2026-04-02

GitHub pushed org-level custom instructions across the Copilot stack.

3. Claude Code auto mode: a safer way to skip permissions

• Source: Anthropic Engineering
• Date: 2026-03-25

Anthropic replaced manual permissions with input probes and output classifiers.

Signal Technique

• Name: Treat the surrounding system as first-class
• Why it matters: The reusable value is in eval, policy, classifier, and harness.

Observations

• The strongest signal is measurability, configurability, and auditability.