代理框架
Hermes Agent 0.20.5 Adds Keyless Search Across Five Providers and Per-Task Reasoning Effort for Scheduled Jobs
Nous Research has rolled keyless web search, persistent memory for scheduled tasks, and execution-stall safeguards into Hermes Agent 0.20.5. Although labeled a patch release, it merges roughly 323 PRs; full migration notes will not arrive until 0.21.0.
Nous Research tagged [Hermes Agent 0.20.5](https://github.com/NousResearch/hermes-agent/releases/tag/v2026.8.19) on August 21, consolidating roughly 323 PRs and 746 commits since 0.20.4 into a stable tag for Docker images, hosted deployments, and new installations. The date in the version number and the August 19 date shown on the release page do not match the actual tagging time, so deployment systems should pin Git tag `v2026.8.19` and commit `fcbd107` rather than relying solely on the displayed date.
The most practical change for agent developers is that new installations can now enable web search without a search API key, using five free providers with rotation and circular failover to maintain availability. This lowers the barrier to starting a local agent, but the rate limits, result quality, and privacy policies of free backends are not equivalent to those of controlled enterprise search services. Sensitive queries should still use an explicitly configured provider.
The scheduling subsystem now preserves memory across runs and allows each job to set a fixed reasoning effort from `none` to `ultra`. The official [scheduling documentation](https://hermes-agent.nousresearch.com/docs/user-guide/features/cron) states that model and reasoning settings can be changed only by the user or an administrative interface; the agent cannot raise them on its own. If a model does not support the specified level, the system caps it at the highest supported level. Scheduled jobs can also run a script first and use `wakeAgent: false` to skip the LLM call when no new data is available, which is especially useful for controlling the cost of high-frequency monitoring.
The release also adds tool-loop stall protection, update receipts, `--plan` validation for fleet updates, and commands for listing and cleaning up worktrees. However, the maintainers have explicitly said that comprehensive, consolidated release notes will not be available until 0.21.0. For now, teams should still review the differences between 0.20.4 and 0.20.5 before upgrading. Hermes' local execution backend is not isolated by default, so engineering teams allowing agents to process untrusted content should prioritize Docker, a cloud sandbox, or restricted network egress.