Agenthood Release Digest: Skills, Security, and Patches — August 16, 2026

Date: August 16, 2026

Author: Agenthood Team

We've shipped four releases over the past two days. v3.35.0 expands the skills library with patterns and members from Copilot and deepagents, v3.34.0 hardens trust boundaries across the agent core, and patches v3.34.1 and v3.33.1 resolve follow-up issues from our internal reviews.

v3.35.0 — Skills library expansion

v3.35.0 brings the skills library closer to parity with our Copilot and deepagents heritage. Four Copilot patterns — review, readme, api-doc, and unit-test — are now merged directly into existing members, so you get these behaviors without adding extra members. We've also ported ten skills from the Copilot library and one from deepagents.

These additions give agents more specialized capabilities for code review, planning, debugging, accessibility auditing, and long-term memory.

v3.34.0 — Trust and security hardening

This release closes several trust-boundary gaps found in our security review. The common thread is that agents now treat external context as untrusted and fail closed when boundaries are ambiguous.

Trust-boundary hardening

• Untrusted context: Retrieved and project context is now delimited and escaped before entering prompts, closing injection paths.
• Redaction: Traces are redacted by default, redaction fails closed, and Sentry payloads are redacted before sending.
• Delegation: Delegation is opt-in and restricted to read-only roles; restricted members cannot delegate, and empty tool sets fail closed.
• Review gate: Ambiguous verdicts now fail the gate rather than pass.
• ReActLoop: Iterations are bounded by a maxSteps guard to prevent runaway loops.
• Decision IDs: Now use higher entropy to avoid collisions.

Alignment and consistency

• Tool grants: Member tool grants align with permission docs, closing delegation boundaries.
• Constructors: Unified across agents, with run errors delegated to the CLI.
• Memory: residualMemory is forwarded through agent option bags, and docs are aligned with memory.write scoping.
• DeveloperAgent: Style and prompt boundaries are aligned with the rest of the agent set.

Tooling and observability

• CI scripts: The decision gate and test-changed script are hardened against review findings; flag-like test paths and signal-death exits are handled correctly.
• Scripts: The academy site title is escaped.
• Observability: Background failures remain visible in the console even without a DSN configured.

v3.34.1 and v3.33.1 — Patch fixes

• v3.34.1: Resolves auditor follow-up warnings for scripts and delegation.
• v3.33.1: Restores Oracle model attribution on failures after the runWithExecutor refactor, closing issue #435.

What's Next

We're continuing to merge and port skills from other libraries, and the security review work in v3.34.0 will inform further hardening of trust boundaries and delegation. Expect more incremental improvements in the coming weeks.