Skip to main content

Agenthood 3.37.1, 3.37.0, and 3.36.0: Mind Virus Defense and Delegation Hardening

Date: August 21, 2026

Author: Agenthood Team

We've shipped three releases over the past two days. The headline work is a completed mind virus defense across the runtime, plus several delegation and propagation fixes. We've also added a conversational style section to every Society member's skills.

v3.36.0 — Mind Virus Defense and Propagation Hardening

This release completes the security work started in earlier patches. We've closed several gaps around message propagation, delegation nesting, and prompt injection.

Security and Runtime

  • Runtime split: The runtime is now fully split, with flattened delegation nesting and an escaped user_query boundary. Delegated tasks no longer inherit propagation context that could be abused.
  • Mind virus immunity warning: System prompts now include a warning about mind virus immunity. This gives agents an explicit guardrail against persona hijacking attempts.
  • SKILL.md integrity check: Skills are now verified at injection time. If a skill file has been tampered with, the runtime rejects it before it reaches an agent.
  • Corrupt lockfile handling: The runtime now recovers gracefully from corrupt lockfiles instead of failing silently or propagating bad state.
  • Marker clamp: Propagation markers are clamped to safe bounds, preventing runaway marker values.
  • Redaction default: Redaction is now on by default for sensitive fields, reducing the chance of accidental leakage.

Propagation and Delegation

  • Distinct markers required: Propagation now requires distinct markers before it fires. This prevents accidental propagation loops when markers collide.
  • No propagation in delegated tasks: The tools module now forbids message propagation inside delegated tasks. Delegation is strictly one-way.
  • Anomaly signals: We've added viral-persona and propagation anomaly signals to the core. These surface suspicious behavior early, making it easier to spot an infection in progress.

Test Coverage

  • 461 fixes covered: The release includes test coverage for all fixes related to issue #461, so the defense is verified end to end.

v3.37.0 — Conversational Style for Society Members

A smaller release, but a useful one for anyone building with Society agents.

  • Conversational style section: All Society members now have a conversational style section in their skills. This gives each agent a consistent, documented voice, which makes multi-agent conversations more predictable and easier to tune.

v3.37.1 — Delegation Cleanup

A quick patch release following up on the 3.36.0 work.

  • MemberAgent delegation smell: We cleaned up a code smell in MemberAgent delegation. The delegation path is now simpler and less error-prone.
  • The-mailman sanitized: The internal the-mailman component has been sanitized. This removes a potential source of malformed messages in high-volume mail flows.

What's Next

The mind virus defense work in 3.36.0 is the kind of foundational hardening that usually precedes more visible features. We expect the next releases to build on the anomaly signals and integrity checks, likely with more observability tooling around propagation and delegation. The conversational style work in 3.37.0 also opens the door for more personality-driven agent configurations in future Society releases.