Hermes v0.17 → v0.20.4 upgrade briefing

Last reviewed: August 2026. Built from official Hermes release notes, the current Hermes docs, and the three Clay-specific migration commits carried forward after the v0.20.4 tag.

Morning verdict

This is less a cosmetic update and more a workflow upgrade. Hermes is now better at staying on task, farming work out to background agents, routing itself across channels, recovering from tool failures, and giving you surfaces to manage memory, agents, bots, cron, Desktop, MCP, and gateway behavior.

Scan tags: worth reviewing · subagents · MoA · verified completion · Desktop + Bot Mode · gateway reliability · custom Clay migration

Read this first

1. Hermes is more autonomous, but with better finish discipline. /goal gained stronger completion contracts and evidence-oriented workflows, so long tasks can be judged against proof instead of vibes. v0.18.0
2. MoA is now a normal model-selection surface. Mixture of Agents presets appear as selectable moa provider models; advisors run first and the aggregator remains the acting model. v0.18.0
3. Background subagents matter more. delegate_task can run batched scouts in parallel, return immediately, and post results back later. Delegation docs
4. Desktop and gateway are more central. The v0.20 line adds more Desktop, connection, Bot Mode, gateway, cron, and messaging reliability work. v0.20.4
5. Your local upgrade includes custom frosting. Photon batching, Telegram reminder buttons, and channel-specific model overrides were ported after the upstream v0.20.4 tag, so treat them as Clay-specific migration features, not official release features.

Version-window timeline

Biggest workflow changes

1. Fan-out research is now a default move

Use subagents as cheap scouts, not just “helpers.” Good pattern:

Spin up 3 scouts: release notes, current docs, and repo/runtime impact.

Each returns source-backed deltas and caveats. Reconcile the results into one verified artifact.

Practical notes from the current docs:

• Batched delegatetask(tasks=[...]) runs subagents in parallel, capped by delegation.maxconcurrent_children.
• Top-level delegation returns a background handle immediately; completed results post back later.
• Subagents inherit the parent’s enabled toolsets; they cannot widen capabilities per task.
• Leaf subagents cannot call delegatetask, clarify, memory, sendmessage, or cronjob.
• For repo-editing parallel work, consider delegation.worktree_isolation: true.
• Long-lived/durable work still belongs in cron or a tracked background process, not delegation. Delegation docs

2. MoA is for judgment moments

MoA is a virtual provider/model path, not a toolset. Use it when multiple perspectives are worth the latency/cost: hard synthesis, code review, decision memos, contentious research, or final critique.

Key gotchas:

• Advisors run in parallel, but the turn waits for the slowest advisor.
• The aggregator is the acting model: final answer and tool calls come from it.
• /moa <prompt> is one-shot sugar; it does not permanently switch the session model.
• Keep reference output short with referencemaxtokens.
• Default fanout is usually enough; per_iteration is fresher but much more expensive.
• Enable moa.privacy_filter if reference models may echo sensitive context. MoA docs

3. “Done” should mean evidence

The biggest behavioral upgrade is not a new button. It is a stronger expectation that tasks finish with proof.

For your work, ask Hermes to read back the target before it reports success:

• Web: fetch the live URL and verify a distinctive string.
• GitHub: show branch, commit, PR URL, and CI state.
• Cron: list the job, inspect last_status / delivery target, and verify important deliveries.
• Trading or external APIs: surface the real order/result status, not just the signal.
• Desktop/browser: capture fresh state after input.

This matches the v0.18 /goal completion-contract direction and your existing preference for verified outcomes. v0.18.0

4. Desktop is becoming the operator console

Desktop is no longer just “chat with a nicer skin.” The current docs describe Desktop using the same Hermes core, config, sessions, skills, memory, and API keys as CLI/gateway. It adds visible management surfaces for models, tools, MCP, gateway connections, sessions, context usage, memory graph / learning journey, and Bot Mode. Desktop docs

For Clay: use Desktop for morning review, session/bot navigation, and visual inspection. Use CLI/repo tools for exact commands and deep development.

5. Gateway/channel routing is now infrastructure

The gateway now supports a broad platform set and more per-channel behavior. The current docs emphasize platform setup, platform-specific capability tables, channel overrides, adapter pause/resume, and circuit-breaker behavior. Messaging docs

Practical moves:

• Keep channel-specific model/system overrides explicit and documented.
• Use Telegram topics/buttons for control surfaces, not only notifications.
• Treat Photon/iMessage as a phone-native intake surface; test batching and authorization.
• Verify exact delivery targets for scheduled jobs and reminders.

6. Cron can be more stateful and less noisy

Cron now has stronger configuration surfaces: attached skills, provider/model override, pre-run scripts, contextfrom, workdir, platform delivery, continuity, and script-only noagent jobs. Cron docs

Clay defaults worth using:

• Use America/Chicago, not fixed offsets.
• For “today” logic, determine live local date/time at run start.
• Use [SILENT] for intended no-op watcher runs.
• Use script-only no_agent=True for deterministic watchdogs and health checks.
• Do not trust “cron run succeeded” as proof of visible delivery; verify the target when debugging.

7. Memory, skills, and curator need hygiene

The built-in memory remains compact. Richer project knowledge belongs in skills, session search, local Markdown hubs, docs, or external memory systems. Memory docs, Skills docs

Post-upgrade options:

• Enable memory.write_approval: true if you want tighter control over always-injected memory.
• Enable skills.write_approval: true if you want diff review before new/changed skills land.
• Use /journey / learning views to inspect what Hermes learned.
• Let curator manage agent-created skills, but remember it does not delete bundled/hub skills automatically. Curator docs

8. Plugins and MCP are the right home for durable integrations

If an integration is reusable, do not bury it in a prompt. Use:

• Skill for procedure.
• Plugin for packaged Hermes capability. Plugin docs
• MCP server for external tools/resources that should be discovered and called through the standard protocol. MCP docs

MCP got especially important in this window: tool naming, health checks, server filtering, OAuth, Hermes-as-MCP-server, and v0.20.3’s SDK/protocol migration all make it more real as an integration layer.

Custom migrated features

These are part of your migrated Hermes, not official upstream v0.20.4 release notes.

Suggested next experiments

1. Cheap delegation preset — 20 minutes

Set a cheaper delegation.model and try a 3-scout research task.

1. MoA preset trio — 45 minutes

Create presets for research-synthesis, code-review, and decision-memo; keep reference outputs short.

1. Goal contract template — 30 minutes

Create a standard ask pattern: acceptance criteria, evidence surface, stop condition, and forbidden side effects.

1. Cron audit — 45 minutes

Review timezones, no-op behavior, continuity, delivery targets, and deterministic script-only candidates.

1. Desktop morning console — 30 minutes

Open Desktop and inspect sessions, bots, memory/learning, gateway connections, and active work.

1. Phone control surface test — 30 minutes

Test Photon batching and Telegram reminder buttons in the real phone workflows you actually use.

1. MCP/plugin inventory — 30 minutes

List enabled MCP servers/plugins, disable broad surfaces you do not use, and document the ones that matter.

Post-upgrade hygiene commands

hermes update --check

hermes config check

hermes config migrate

hermes skills check

hermes mcp list

hermes doctor

Use hermes gateway status and hermes cron list for the two parts most likely to affect your daily remote workflow.

Source map

Official release notes

• v0.17.0 — The Reach Release: https://github.com/NousResearch/hermes-agent/releases/tag/v2026.6.19
• v0.18.0 — The Judgment Release: https://github.com/NousResearch/hermes-agent/releases/tag/v2026.7.1
• v0.19.0 — The Quicksilver Release: https://github.com/NousResearch/hermes-agent/releases/tag/v2026.7.20
• v0.20.0 — The Herald Release: https://github.com/NousResearch/hermes-agent/releases/tag/v2026.8.3
• v0.20.1: https://github.com/NousResearch/hermes-agent/releases/tag/v2026.8.13
• v0.20.2: https://github.com/NousResearch/hermes-agent/releases/tag/v2026.8.16
• v0.20.3: https://github.com/NousResearch/hermes-agent/releases/tag/v2026.8.16.2
• v0.20.4: https://github.com/NousResearch/hermes-agent/releases/tag/v2026.8.18

Official docs

• Docs home: https://hermes-agent.nousresearch.com/docs/
• Updating: https://hermes-agent.nousresearch.com/docs/getting-started/updating
• CLI commands: https://hermes-agent.nousresearch.com/docs/reference/cli-commands
• Slash commands: https://hermes-agent.nousresearch.com/docs/reference/slash-commands
• Configuration: https://hermes-agent.nousresearch.com/docs/user-guide/configuration
• Delegation: https://hermes-agent.nousresearch.com/docs/user-guide/features/delegation
• MoA: https://hermes-agent.nousresearch.com/docs/user-guide/features/mixture-of-agents
• Cron: https://hermes-agent.nousresearch.com/docs/user-guide/features/cron
• Memory: https://hermes-agent.nousresearch.com/docs/user-guide/features/memory
• Skills: https://hermes-agent.nousresearch.com/docs/user-guide/features/skills
• Curator: https://hermes-agent.nousresearch.com/docs/user-guide/features/curator
• Plugins: https://hermes-agent.nousresearch.com/docs/developer-guide/plugins
• MCP: https://hermes-agent.nousresearch.com/docs/user-guide/features/mcp
• Messaging/gateway: https://hermes-agent.nousresearch.com/docs/user-guide/messaging
• Desktop: https://hermes-agent.nousresearch.com/docs/user-guide/desktop
• Voice: https://hermes-agent.nousresearch.com/docs/user-guide/features/voice-mode