Two large agent-platform releases landed at almost the same time, but they are not trying to win in exactly the same way.
OpenClaw 2.0 is the OpenClaw 2026.8.1 release. It pushes OpenClaw toward a broader control plane for sessions, paired devices, cloud workers, permissions, interactive results, credentials, memory, and recoverable operations.
The new official Hermes release is Hermes Agent v0.21.0, not a product formally named "Hermes 2". Nous Research calls it the Pantheon release. Its center of gravity is a visible society of named agents: group chats, bot-to-bot messages, scheduled agents that remember previous runs, live subagent steering, an MCP command center, and an in-app browser the agent can operate.
That distinction is more useful than asking which product has the longer feature list. OpenClaw 2 is moving toward distributed agent operations. Hermes v0.21 is moving toward an operator-facing multi-agent workbench. Both overlap, and both now contain serious security and reliability controls. The right choice depends on how your team wants work to move, where it runs, and who approves consequential actions.
The answer in one table
| Decision area | OpenClaw 2.0 | Hermes Agent v0.21.0 |
|---|---|---|
| Official release | OpenClaw 2026.8.1 | Hermes Agent v0.21.0, tag v2026.8.31 |
| Primary idea | Distributed agent control plane | Visible multi-agent workbench |
| Multi-agent work | Durable progress visibility across subagents and accumulating edits | Named bots, group chats, peer messages, live steering and stopping |
| Where work runs | Gateway, paired devices and cloud workers with reusable workspaces | CLI, Desktop and Gateway profiles, with agent communication across profiles and gateways |
| Scheduled work | Exact-operation recurring permissions and conversation-bound automation controls | Cron memory, continuity, durable notepads and change-aware monitor mode |
| Operator interface | Structured questions, session progress cards, widgets and dashboards | Bot Mode, MCP command center, richer CLI status and controllable in-app browser |
| Credential approach | Masked credential requests and optional protected-secret substitution proxy | Redaction hardening, protected instruction files and approval coverage for destructive commands |
| Upgrade risk | Higher for affected OpenProse, OpenAI route and external plugin configurations | Broad release surface; use the update plan, config migration and installation-specific route |
| Recommended first evaluation | One distributed or permission-sensitive workflow | One multi-agent research, monitoring or engineering workflow |
This is an architectural comparison, not a benchmark. The official release notes do not establish that either system is faster, cheaper, more accurate, or more secure in every deployment.
What changed in OpenClaw 2.0
OpenClaw 2026.8.1 expands the system around the model. That matters because most agent failures in production are not caused by a weak chat response. They happen at the boundaries: state, permissions, credentials, tool execution, retries, updates, and unclear operator control.
Sessions can leave the Gateway
OpenClaw can run sessions on paired devices or cloud workers, move the workspace with the session, and reuse warm machines or project seeds later. This is a meaningful shift for teams that do not want every long-running job tied to one Gateway machine.
The release also adds durable progress cards and more visibility into subagent activity and edits. Operators can reload the interface without losing the current view of what is happening. That does not prove the work is correct, but it makes long-running work easier to inspect.
Questions and results become structured interfaces
Agents can ask structured questions through web and native cards, messaging buttons, or plain text. They can also return interactive widgets, pin them to session dashboards, receive permission for specific actions or network origins, and export rendered views as images.
This is more than UI polish. A structured question can reduce ambiguity before an action, and a constrained widget can expose a narrow decision instead of handing the operator a wall of generated prose.
Credentials and recurring permissions get narrower controls
OpenClaw 2 lets an agent request a credential through a masked prompt without placing the value in chat or model context. An optional proxy can restrict protected-secret substitution to approved destinations.
Recurring automation can receive permission for an exact operation. The operator can inspect or revoke that permission, and a changed job or operation requires fresh approval. This is a better control model than giving a scheduled agent a broad standing approval, but it still depends on correct configuration and a trustworthy host.
Memory becomes more active
The release adds personal conversation recall under defined conditions, background memory consolidation, and automatic capture of reusable lessons or skills with controls. These features may reduce repeated setup, but they also increase the importance of source boundaries, retention policy, inspection, and the ability to disable unwanted learning.
Teams should not enable active memory simply because it exists. Decide which conversations may be recalled, which sources can become durable memory, and how a person can inspect or remove bad material.
The upgrade has real breaking changes
OpenClaw 2026.8.1 removes the bundled OpenProse plugin and /prose command. The release notes direct affected users to run openclaw doctor --fix and migrate to the upstream Agent Skill while retaining existing .prose source files.
It also migrates shipped codex/* and openai-codex/* references to openai/*. External plugin authors have additional SDK migrations to prepare for the documented removal gates.
This is not a blind auto-update release for a customized production install. Back up configuration and state, review the breaking sections, run the doctor, restart the Gateway, and test the workflows that use affected providers or plugins.
What changed in Hermes Agent v0.21.0
Hermes v0.21 turns multi-agent work into something the operator can see and direct. The release is large, but six changes define its operating model.
Bot Mode makes agents visible as a team
Bot Mode is bundled into the desktop app. Agent profiles get names and deterministic avatars, and several bots can participate in a shared room with the operator. A user can mention a specific bot instead of treating every agent as an invisible background process.
The useful part is not the avatar. It is the durable social surface. Team members can inspect who said what, where a handoff occurred, and which conversation owns the next step.
hermes peer creates durable bot-to-bot messages
Agents can message one another by handle across profiles and gateways. Replies land in each agent's canonical Bot Chat, so a research agent can hand findings to a coding agent without the exchange disappearing into an opaque internal trace.
Durability improves inspection, but it does not remove coordination risk. Teams still need clear ownership, bounded permissions, and a rule for resolving conflicting agent recommendations.
Scheduled agents remember previous runs
Hermes cron jobs can load and update persistent memory. With continuity enabled, one run carries its output into the next. Jobs also receive durable notepads, and monitor mode can skip the model when nothing has changed.
This is a practical improvement for recurring monitoring. A job can avoid repeating yesterday's alert and preserve its working context. The operator should still define the source, change detector, deduplication rule, failure notification, and retention period.
Subagents can be steered while they run
The delegate_task flow can list running children, steer one with a course correction, or stop it while preserving partial output. The release also supports optional JSON Schema validation for child output and surfaces per-delegation cost in the result.
That moves delegation away from a launch-and-wait pattern. It is especially relevant when a task is expensive, long-running, or likely to discover evidence that changes the plan.
MCP becomes an operator dashboard
The desktop MCP page combines server management and the catalog. It adds imports, background health checks, reauthentication warnings, usage information, and explicit confirmation for hermes:// installation links.
This can make a large tool fleet easier to operate. It does not make every MCP server safe. Each server still needs a source review, permission boundary, credential scope, and removal path.
The desktop browser becomes actionable
Hermes can navigate, click, and read the browser inside the desktop app. This supports documentation research and web application debugging without switching to a separate automation surface.
Browser control also expands the failure surface. Authentication state, destructive clicks, file downloads, cross-origin content, prompt injection, and form submissions require clear approval rules and test environments.
Security: the important differences
Both releases treat security as an operating-system problem, not a safety paragraph.
OpenClaw 2 emphasizes protected credential flow, exact-operation automation grants, model allowlists, plugin trust and provenance, recoverable backups, security audits, and migration tooling. Its documentation is explicit that shared inbox or team hardening is not hostile-tenant isolation. Separate gateways, operating-system users, or hosts are needed when operators do not share a trust boundary.
Hermes v0.21 protects agent instruction files such as AGENTS.md, skills, and memory behind write approval. It expands redaction across environment reads, terminal errors, checkpoints, and logs; adds destructive Windows commands to approval coverage; and adds security scanning for plugin installation. The release also records a supply-chain response that removed a compromised catalog entry and skill.
Neither release turns an agent host into a zero-trust platform automatically. The deployment owner remains responsible for host security, credential issuance, network exposure, backups, audit retention, and the permissions granted to every connected tool.
OpenClaw 2 or Hermes: choose by constraint
Choose OpenClaw 2 for distributed and permission-sensitive operations
OpenClaw deserves the first pilot when the deciding requirement is one of these:
- work must move between a Gateway, paired devices, or cloud workers;
- reusable workspaces and warm worker environments matter;
- recurring operations need an exact approval that becomes invalid when the job changes;
- protected credentials must stay out of chat and model context;
- interactive session widgets or structured approval surfaces are central;
- the team needs a documented import path from another agent system.
Choose Hermes v0.21 for visible multi-agent coordination
Hermes deserves the first pilot when the deciding requirement is one of these:
- operators want named agents and shared group conversations;
- bot-to-bot handoffs must remain durable and inspectable;
- recurring monitoring needs memory, continuity, notepads, and no-change suppression;
- long-running subagents need live steering or early stopping;
- a large MCP fleet needs one desktop management surface;
- browser research and debugging should happen inside the same desktop environment.
Do not choose from the feature count
A long release note is not evidence that a product fits your workflow. Start with the bottleneck, the systems involved, the approval boundary, and the observable completion condition. Then choose the smallest architecture that can pass a representative evaluation.
A safe upgrade plan for OpenClaw 2
- Record the current version, profiles, Gateway configuration, plugins, providers, channels, automations, and model routes.
- Stop the Gateway and create a verified backup of state and workspaces.
- Read the OpenProse, OpenAI route, provider package, and plugin SDK migration sections that apply to the installation.
- Upgrade in a staging or recoverable environment.
- Run
openclaw doctor --fix, review every reported migration, restart the Gateway, and runopenclaw status. - Run
openclaw security auditand use the deep mode only when live Gateway and plugin collectors are appropriate. - Test one normal workflow, one permission denial, one credential request, one scheduled job, and one recovery path.
- Keep the previous verified archive until the new installation has completed those checks.
OpenClaw's restore flow stages an archive rather than activating it in place. The migration guide also warns that older channel state can desynchronize credentials and that plugin dependencies may need reinstalling.
A safe upgrade plan for Hermes Agent v0.21
- Export or back up the active Hermes profile, configuration, memory, skills, sessions, and installation-specific state.
- Run the update check or plan before changing the installation.
- Use
hermes updatefor a supported source installation. Use the image or package manager route when the install is managed by Docker, pipx, Homebrew, or another external mechanism. - Review configuration migrations and new defaults instead of accepting every prompt automatically.
- Confirm the Desktop or Gateway starts, profiles load, MCP servers authenticate, and scheduled jobs retain the intended continuity.
- Test agent-to-agent messaging, one steered subagent, one stopped subagent with partial output, and one browser action in a safe environment.
- Inspect approvals and redaction with a deliberately denied command and a non-secret test value.
- Keep a compatible rollback package and configuration backup until the representative workflow passes.
Hermes documentation warns that rolling back can expose configuration incompatibilities. Run the configuration check after a rollback rather than assuming an older binary understands newer settings.
The business pilot both systems should pass
Use the same pilot brief for both products so the comparison is fair.
Define one job
Choose one recurring workflow with a real owner. Good candidates include research handoff, issue triage, controlled content review, documentation maintenance, or change-aware monitoring. Avoid a first pilot that can send money, publish publicly, delete records, or contact customers without approval.
Draw the authority map
List every system the agent can read, draft to, and write to. Mark every action that requires a person. Name the credential used, its scope, and how it can be revoked without disabling unrelated work.
Build the evaluation set
Include normal inputs, missing context, conflicting sources, a tool outage, a prompt-injection attempt, an approval denial, and a partial failure after one successful action. Measure task completion, unsupported claims, unnecessary actions, correct escalation, and recovery.
Define the stop condition
Stop the pilot if the system cannot show what it did, cannot respect a denied action, loses required state, or needs broader permissions than the business can justify. Expand only after the bounded workflow passes.
For a broader architecture primer, read AI agent vs chatbot for business. If the first use case is customer-facing conversation rather than multi-step work, compare the narrower AI business chatbot route. Teams that need tool-connected operational work can review YAG's AI agents service.
Our verdict
OpenClaw 2.0 is the more compelling update for teams that see the agent platform as a distributed control plane. Its worker portability, exact-operation permissions, structured interfaces, credential flow, migration tooling, and active memory controls point in that direction.
Hermes Agent v0.21.0 is the more compelling update for teams that want multi-agent work to feel visible and steerable. Bot Mode, durable peer messages, cron continuity, live subagent control, MCP management, and an actionable desktop browser make that operating model concrete.
There is no honest universal winner. Run the same bounded workflow on both systems, keep permissions narrow, and judge the complete operation: setup, intervention, failure handling, auditability, recovery, and maintenance.
Send YAG the workflow you want to evaluate. We will map the systems, authority, test cases, and smallest defensible pilot before recommending an agent platform.
Official sources
- OpenClaw: 2026.8.1 release notes
Official release documentation - OpenClaw: signed GitHub release v2026.8.1
Source tag and release record - OpenClaw: migration guide
Backup, restore and verification steps - OpenClaw: security audit CLI
Audit modes and safe fixes - Nous Research: Hermes Agent v0.21.0 release
Official Pantheon release record - Hermes Agent: official updating guide
Update, migration and rollback instructions
Frequently asked questions
Is OpenClaw 2 a new AI model?
No. OpenClaw 2 is the name used for the OpenClaw 2026.8.1 platform release. It expands the agent runtime, session, worker, automation, permission, widget, memory, migration, and security layers. The model remains a provider choice inside the system.
Is Hermes 2 the official name of the new Hermes Agent release?
No. Nous Research labels the current release Hermes Agent v0.21.0, dated August 31, 2026. This comparison uses "Hermes 2" only as the informal wording people may use when referring to that update.
What is the biggest difference between OpenClaw 2 and Hermes Agent v0.21?
OpenClaw 2.0 puts more emphasis on a distributed agent control plane with paired devices, cloud workers, scoped automation permissions, widgets, credential handling, and migration tooling. Hermes Agent v0.21 centers the operator experience around named agents, group chats, bot-to-bot messages, persistent cron continuity, live subagent steering, an MCP command center, and a desktop browser the agent can drive.
Which system is better for a small business?
Neither is a universal winner. OpenClaw is the stronger candidate when distributed sessions, explicit operation permissions, worker portability, and control-plane features are the deciding constraints. Hermes is the stronger candidate when the team wants a visible multi-agent workspace, durable bot conversations, scheduled continuity, and direct operator steering. Test one bounded workflow before granting production write access.
Can a company migrate from Hermes to OpenClaw?
OpenClaw's official migration guide includes a Hermes import path for supported configuration, providers, MCP servers, memory, skills, and selected environment keys. Preview the migration, make a verified backup, stage the result, and inspect unsupported or conflicting items before activation.
Should an existing installation upgrade immediately?
Not without checking the release notes and rollback path. OpenClaw 2026.8.1 includes breaking OpenProse and OpenAI route migrations. Hermes provides update checks, plans, configuration migration, and separate instructions for package-managed or container-managed installs. Back up state and verify a representative workflow after upgrading.