← All agent and harness briefs
Agent runtimeautonomousopscodingtestedunattended capable

Operational brief

Hermes Agent

Nous Research agent runtime with skills, memory, and multi-surface gateways.

Official Hermes Agent image from Nous Research

Choose it when

You want reusable skills and memory across chat surfaces, plus scheduled work whose tools, outputs and failures you own.

Skip it when

You want only an IDE coding assistant, or can't assign an owner to unattended jobs and their failures.

01 / Classification

Autonomy evidence

This is our classification, not a vendor badge. The record says what the product normally does. The gaps stay visible.

How we classify it
Unattended scheduler/heartbeat with write-tools; failure has a real owner.
Runs unattended
Yes — that capability is documented in this brief.
Trigger
Human CLI/chat requests or configured schedules; cron execution needs the gateway scheduler running.
State
Profile-scoped memory files persist across sessions; job definitions and run outputs live on disk. Scheduled runs start fresh sessions rather than inheriting a chat implicitly.
Tools
Configured terminal and file tools can write within their granted host/backend authority. Cron toolsets are configurable, not automatically every available tool.
Human control
The operator owns schedules, toolsets and delivery targets, can interrupt interactive work, and can pause or remove jobs. This catalog retains the established desk-run owner scope, not a reliability guarantee for arbitrary tasks.

02 / Operations

Control reality

Useful agents eventually touch something real. This is the operational picture we have, not a promise hidden in a sales page.

Where it runs
Self-run CLI and gateway; terminal backend choice determines where commands execute.
Data boundary
Local memory and job/output files aren't a no-egress promise. Remote models, external tools and messaging delivery can receive selected context or results; retention depends on those services and your configuration.
Identity & permissions
The operator selects model/tool providers and messaging accounts. Profile-scoped memory helps separate agents; it isn't a blanket process or filesystem sandbox.
What it can write
host filesystem · shell · scheduled jobs
Surfaces
CLI · Telegram / Discord / Slack · Desktop · Cron / gateway
Export / lock-in
Memory files and plain-JSON jobs are inspectable local state; cron outputs are Markdown files. Backing up those files isn't a certified migration of every provider or channel setting.
Operational limits
Tested here means the established desk-run cron, write-tool and owner scope only. This refresh reviews documentation, not the latest release in operation. Check features against your installed build; remote model and tool costs depend on the services you choose.
Watch for
Check your installed build separately from upstream. Review each job's toolset, working directory and delivery target; a self-hosted runtime doesn't keep remote-model context on the host. Upstream v0.22.2 corrects v0.22.1's wrong-commit build; its known issue is that sessions run under the 0.22.1 gateway can't be deleted in 0.22.2. The release says main-only work ships with v0.23.0, so don't treat every current README capability as stable v0.22.2 behavior.

03 / Receipts

Evidence ledger

Here is what this brief is built on. “Documented” means we found a source. “Tested” means this desk actually ran it. Those are different things.

Evidence status
tested
Source set
7 primary sources
Evidence review
2026-10-17

Corrections

  • 2026-10-13 · Removed the dated provider/auth snapshot from watch_for. Provider availability is deployment-specific, not a durable product capability claim.
  • 2026-10-13 · Replaced a desk-payload date used as release metadata with upstream v0.21.7's API publication timestamp. Existing tested status remains limited to established desk-run operations, not that upstream release or all current docs.
  • 2026-10-17 · Refreshed upstream release metadata from v0.21.7 to v0.22.2 using the release API publication timestamp. The release corrects v0.22.1's wrong-commit build, retains a known session-deletion issue for sessions run under the 0.22.1 gateway, and separates main-only v0.23.0 work. This is not an installed upgrade or an expansion of the established desk-run tested scope; the upstream release is not runtime-tested here.

Technical dossier

Technical review

Pillar weights describe the proposed rubric, not measured results. No overall score or security certification has been assigned.

Desk use
We use Hermes for desk operations.
Potential conflict
Our reliance on Hermes creates a potential familiarity and operational-interest bias; it must not be treated as independent validation or a favorable score.
Audit status
This review is not an independent security audit, certification or executed standardized lab review.
Observation boundary
Operational use is the disclosed desk observation; the technical summaries are vendor-documentation findings, not runtime-tested in this review. Every pillar score, overall score and security tier remains null/unassessed.
Vendor documentation
Summaries use the retained upstream README and SECURITY.md, not an implementation audit, installed-version inspection or newly executed test.
Desk observation
The desk uses Hermes operationally; that use is disclosed, not evidence that this pilot protocol passed or that current upstream security behavior matches the installed build.
Supplement scope
Supplement only: retain the existing decision, autonomy, control and source sections unchanged; the old tested label does not supply standardized lab results.
Historical release metadata
Upstream v0.22.2; release API publication, not the desk's installed version · 2026-10-15T08:21:32Z
Legacy tested scope
Tested here means the established desk-run cron, write-tool and owner scope only. This refresh reviews documentation, not the latest release in operation. Check features against your installed build; remote model and tool costs depend on the services you choose.
Version and operational caveats
Check your installed build separately from upstream. Review each job's toolset, working directory and delivery target; a self-hosted runtime doesn't keep remote-model context on the host. Upstream v0.22.2 corrects v0.22.1's wrong-commit build; its known issue is that sessions run under the 0.22.1 gateway can't be deleted in 0.22.2. The release says main-only work ships with v0.23.0, so don't treat every current README capability as stable v0.22.2 behavior.
Protocol status
Multi-hop: Not tested. Recovery: Not tested. Canary: Not tested. Metrics: Not measured.
Limitations
No latency, cost, recovery, prompt-injection or canary measurements; no ranking, security certification or formal audit.
Limitations
Installed version, active terminal backend, whole-process wrapper, mounts, egress policy and enforcement were not inspected in this pilot.
Limitations
Upstream distinguishes OS-level isolation from approval gates, redaction and allowlists; this draft doesn't verify their implementation or guarantee containment.

Documentary claim ledger

Retained source snapshots; capture date and revision not recorded here. Source links point to mutable upstream documents. Quoted documentation is not a semantic audit or proof of installed behavior.

  • https://github.com/NousResearch/hermes-agent/blob/main/SECURITY.md · summaries.architecture
    - **Agent process.** The Python interpreter running Hermes Agent, including any Python modules it has loaded (skills, plugins, hook handlers). - **Terminal backend.** A pluggable execution target for the `terminal()` tool. The default runs commands directly on the host.

    Documented architecture; choosing a boundary is editorial guidance, not a tested configuration.

  • https://github.com/NousResearch/hermes-agent/blob/main/SECURITY.md · summaries.architecture
    What this confines: anything the agent does by issuing shell or file operations. What this does **not** confine: everything the agent does in its own Python process.

    Terminal-backend isolation isn't whole-process containment; installed behavior remains unverified.

  • https://github.com/NousResearch/hermes-agent/blob/main/SECURITY.md · summaries.security
    The default runs commands directly on the host.

    Current upstream default, not an observation of this desk's active backend or installed version.

  • https://github.com/NousResearch/hermes-agent/blob/main/SECURITY.md · summaries.security
    That includes the code-execution tool (spawned as a host subprocess), MCP subprocesses (spawned from the agent's environment), plugin loading, hook dispatch, and skill loading (all imported into the agent interpreter).

    Documented paths outside terminal-backend isolation, not a demonstrated escape or vulnerability finding.

  • https://github.com/NousResearch/hermes-agent/blob/main/SECURITY.md · summaries.security
    Whole-process wrapping runs the entire agent process tree inside a sandbox. Every code path — shell, code-execution, MCP, file tools, plugins, hooks, skill loading — is subject to the same filesystem, network, process, and (where applicable) inference policy.

    Upstream posture description, not verified enforcement or an implementation guarantee from this desk.

  • https://github.com/NousResearch/hermes-agent/blob/main/SECURITY.md · summaries.security
    This is the supported posture when the agent ingests content from surfaces the operator does not control — the open web, inbound email, multi-user channels, untrusted MCP servers — and for production or shared deployments.

    Upstream recommends whole-process wrapping for this threat model; no configuration change or lab test performed.

  • https://github.com/NousResearch/hermes-agent · summaries.interoperability
    Use any model you want — [Nous Portal](https://portal.nousresearch.com), OpenRouter, OpenAI, your own endpoint, and [many others](https://hermes-agent.nousresearch.com/docs/integrations/providers).

    Documents provider/endpoint choices, not verified universal compatibility, availability or provider terms.

  • https://github.com/NousResearch/hermes-agent · summaries.interoperability
    Hermes has two entry points: start the terminal UI with `hermes`, or run the gateway and talk to it from Telegram, Discord, Slack, WhatsApp, Signal, or Email.

    Documents messaging gateways, not delivery reliability or functioning account integrations at this desk.

  • https://github.com/NousResearch/hermes-agent · summaries.interoperability
    [MCP Integration](https://hermes-agent.nousresearch.com/docs/user-guide/features/mcp) | Connect any MCP server for extended capabilities

    README documents MCP integration; linked guide wasn't fetched, and the vendor's broad wording isn't a tested compatibility guarantee.

  • https://github.com/NousResearch/hermes-agent · summaries.developer_experience
    Full TUI with multiline editing, slash-command autocomplete, conversation history, interrupt-and-redirect, and streaming tool output.

    Documented interface affordances, not a usability study, installation test or developer-experience score.

  • https://github.com/NousResearch/hermes-agent/blob/main/SECURITY.md · lab_evidence.limitations
    Nothing inside the agent process constitutes containment — not the approval gate, not output redaction, not any pattern scanner, not any tool allowlist.

    Documented trust model; neither an audit nor a prompt-injection result, and not evidence that all configurations are safe.

Overall score
Unassessed
Security tier
Unassessed
Architecture and execution state · 25%
Pillar score: Unassessed. Vendor documentation describes a Python agent process with pluggable terminal backends; choose the execution boundary deliberately rather than assuming a remote terminal contains the whole agent.
Security and governance · 25%
Pillar score: Unassessed. Current upstream SECURITY.md says the default terminal backend executes on the host and terminal isolation leaves Python-process paths outside it, so consider whole-process wrapping for untrusted inputs; this policy is not proof of the installed build's isolation.
Protocol/model interoperability · 20%
Pillar score: Unassessed. The README documents selectable model endpoints, messaging gateways and MCP integration, but compatibility across this desk's installed configuration remains unverified.
Empirical benchmarks · 15%
Pillar score: Unassessed. The standardized multi-hop, recovery and canary protocol is untested, so performance, reliability and cost remain unassessed.
Developer experience and observability · 15%
Pillar score: Unassessed. The README documents a terminal UI with history, interrupt-and-redirect and streaming tool output, but setup quality and usability haven't been measured in this pilot.
Laboratory evidence
Laboratory status: Not tested. Not assessed.

04 / Change log

Related coverage

The brief tells you what the product is. Payloads track the changes that may alter the decision.