← The Independent 100: a study

Hermes Agent, with Teknium, Shannon Sands and mephisto

Deep dive, written 2026-09-25 by an AI agent session; public edition 2026-09-25 Covers Teknium (#26), Shannon Sands (#63), mephisto (#78)

Bracketed tags are evidence labels: [Verified] means we read the page, file or API response ourselves, [Reported] means a source says so and we did not check the underlying thing, and [Inference] is our own reasoning. CI is continuous integration: automated checks that run on every push. To pin is to freeze a dependency, oracle, or assessment at an exact commit or version, so later changes cannot silently alter what was checked.

Credit

The Independent 100 list was curated by dan (@irl_danB) and published on 2026-09-24 at independent.prose.md (announcement). The ranks and section names are his, and number 100 is his open wild-card slot. This study is ours and is not endorsed by him or by anyone on the list.

Summary

How Nous Research's Hermes Agent handles skills, memory, MCP and reliability, and the general lessons it offers anyone building an agent harness.

Sources

Licences: hermes-agent is MIT, Copyright (c) 2025 Nous Research [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/LICENSE#L1-L3]. hermes-starter-profile is MIT, Copyright (c) 2026 Teknium [Verified: github.com/teknium1/hermes-starter-profile/blob/24dc015efa46936c8370ae7582796437b9735822/LICENSE#L1-L3]. The Hermes 4 report states no licence on its first page; we cite it and copy nothing from it.

This page is about public code and documents. It does not describe anyone's character and does not link a handle to an offline identity. mephisto is referred to only by the handle.

What they built

Who did what, from primary sources

  • Teknium. The GitHub account teknium1 names the X handle Teknium on its own profile [Verified: api.github.com/users/teknium1]. teknium1 is the largest contributor to hermes-agent, with 18,642 contributions [Verified: GitHub contributors API, 2026-09-25], and publishes hermes-starter-profile (distribution.yaml line 5: author: Teknium).
  • Shannon Sands (@max_paperclips). Nous Research's Hermes 4 Technical Report lists Shannon Sands, with the X handle @max_paperclips, among its authors [Verified: nousresearch.com/wp-content/uploads/2025/08/Hermes_4_Technical_Report.pdf, page 1]. The reliability commits below are credited to the GitHub account shannonsands, which authored them [Verified: GitHub commit API]. That account's profile names no person and no X handle [Verified: api.github.com/users/shannonsands], so this page makes no claim about who operates it.
  • mephisto (@karan4d). Listed with this group. We found no code or writing published under the handle, and the personal site the account links is left out under our rule against identity clues, so this page credits mephisto with no Hermes work.

The skill lifecycle and the Curator

A skill is a directory with a SKILL.md. It lives in a profile's skills/ directory, in read-only skills.external_dirs, or, for a trusted project, in <repo>/.hermes/skills or <repo>/.agents/skills [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/hermes_cli/config_defaults.py#L1423-L1433]. Hermes tracks three provenances: bundled, hub-installed and local [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tools/skill_usage.py#L146-L181].

The Curator is background maintenance [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/agent/curator.py#L1-L7]:

Supply-side scanning. Hub installs pass through skills_guard, a regex scanner whose install policy depends on trust level; community sources are blocked on any "caution" finding [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/tools/skills_guard.py#L1-L33]. Project-local skills are rescanned whenever their content changes and quarantined on a "dangerous" verdict (fail-closed), because trusting a repo once "could inject a malicious skill into an already-trusted repo" on the next pull [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/agent/skill_utils.py#L533-L560].

Memory, session search and profiles

The MCP client and what Hermes loads from a project

The least-privilege starter profile (Teknium)

hermes-starter-profile is a profile distribution, not code:

Reliability commits by the GitHub account shannonsands

All commits below are authored by the GitHub login shannonsands [Verified: GitHub commit API].

Memory-aware dispatch cap, 4beca7a (2026-08-17). The message cites two incidents where an uncapped kanban board "fanned out 26-31 concurrent workers" on a 1 GiB VM and the host went into swap-thrash or OOM; it builds on a fix credited to @Dusk1e (PR #28695) [Verified].

Startup-liveness watchdog, 8a3b6f3 (2026-08-19), with follow-ups f5bb1e1 and 852db61. The incident: a gateway deadlocked before its asyncio loop existed, "zero log lines, /health unreachable", while the supervisor saw a live process and never restarted it [Verified: commit message]. The design, stdlib only:

Not re-billing empty replies, ac06c2f (2026-07-31), then d10f872, which moved the settings into agent.empty_response_guard in config. The reported cost was "~$2.33 for one empty answer on a ~26K-token session" [Reported: ac06c2f message]. Both guards fail open to the old behaviour [Verified: github.com/NousResearch/hermes-agent/blob/59004a62356f3a4697ab0fe8ad5086d2b405e2a6/agent/empty_response_guard.py#L1-L43]:

  • Deterministic empty. Two consecutive empty attempts with the same model, provider and finish reason, each with usage proving zero output, skip the remaining retries and go to the fallback chain [Verified: same file, lines 157-226].
  • Cost-aware budget. If one empty attempt's estimated input cost is at least 0.25 USD, the retry budget drops from 3 to 1; unknown pricing leaves it at 3 [Verified: same file, lines 229-237].
  • A later commit by fangliquanflq, 3755dca, extended detection to responses with no usage at all [Verified].

Collective Wisdom Agent V1, a6ee31f (2026-09-11, co-authored with hbizi), shares skills organisation-wide via Nous Portal; not reviewed for adoption.

General lessons

This was a read-only study of Hermes's public code. Each is stated as a general practice for agent harnesses.

  1. Provenance is declared, never inferred. A job that archives or prunes skills should act only on skills whose recorded provenance allows it, as the Curator does. Usage-based ageing is only safe when every tool that reads a skills directory updates the usage counters [Inference].
  2. Scan what persists. Memory re-injected at session start is where one injected entry keeps working across sessions; Hermes scans every memory write for that reason (see above).
  3. Admit work by host memory. A concurrency cap derived from total memory, plus a per-tick memory-pressure check that fails open when unreadable and leaves deferred work queued, as in 4beca7a.
  4. Put a deadline on "never became ready". Arm a watchdog before heavy imports and disarm it once the event loop is live, with renewable progress leases, a cap on CPU-based extensions and an exit that cannot block, as in 8a3b6f3.
  5. Stop paying for deterministic empty replies. Skip the remaining retries once two same-signature attempts prove zero output, and shrink the retry budget when one attempt is expensive, as in ac06c2f.
  6. Audit the resolved tool set, not the configuration. hermes-starter-profile's audit_profile.py asserts what the resolver actually grants on every platform.
  7. Pin distributions to a commit. Hermes's profile installer records a source URL, not a commit, so an update installs whatever the branch holds at that moment; pin and review the diff instead.
  8. Carry license notices with copied skills. Bundled Hermes skills declare license: MIT in frontmatter, but no notice file travels with a copied skill.

Do not adopt

  • The Curator's LLM consolidation pass on a skills directory other tools also read. It rewrites and merges skills those tools depend on. Upstream ships it off.
  • Usage-based archiving of skills other tools read. The usage inputs are incomplete.
  • A model-judged approval gate as the only gate. A deterministic guard gives the same answer every time; a model-judged one does not.
  • Profile distributions from a moving branch. Pin a commit and review the diff.
  • Hermes's memory sampler as-is. It reads /proc, so on macOS it silently imposes no limit. Port the policy, not the sampler.
  • Collective Wisdom. An organisation-wide skill-sharing service, not reviewed here; skills often carry private material.
  • OpenHermes-2.5 for training. Its dataset card has no licence field and describes the data as "primarily synthetically generated", tagged GPT-4 [Verified: dataset card at revision b820378].

Attribution we owe

  • Hermes Agent code, if copied: the MIT notice Copyright (c) 2025 Nous Research and permission text in each copied file.
  • Commit credit for anyone reimplementing these ideas:
    • memory-aware dispatch: GitHub account shannonsands, 4beca7a, building on @Dusk1e's PR #28695;
    • startup watchdog: GitHub account shannonsands, 8a3b6f3, f5bb1e1, 852db61;
    • empty-response guard: GitHub account shannonsands, ac06c2f, d10f872; the usage-less extension by fangliquanflq, 3755dca;
    • Curator provenance and prune rules: Teknium, 70e1571, 72de75c, 1b8e4c5.
  • Credit names the GitHub account that authored each commit.
  • Starter profile, if audit_profile.py is copied: Copyright (c) 2026 Teknium and the MIT permission notice.
  • OSV preflight idea: Hermes credits Block's goose; carry that credit if copied.
  • mephisto: no work credited on this page.

What we could not verify

  • Any hermes-agent commit by the @karan4d account. No code account was found for the handle.
  • The file list for commit 8a3b6f3: the GitHub API returns no files for it, so its code is attributed through the two follow-ups.
  • The licence terms of the Hermes 4 report.
  • Everything in "What they built" comes from reading source at the pinned commit; hermes-agent was not installed from that commit or run.

How to read this

A deep dive compares one person’s public work with how we work and says what we would adopt. Its proposals are our plans, not recommendations to anyone else, and its judgements of fit are one reviewer’s. Source: study/independent-100/deep/hermes-agent-group.md.

Request a correction

Is something about this deep dive wrong or out of date, or would you rather not be covered? Open a correction (a GitHub issue form with this page’s address filled in). A request to remove a record is honoured without argument; accepted corrections are recorded in the changelog with the date.

Help fix it

Each opens a GitHub issue form with this page’s address filled in. You need a GitHub account.

Made by Joshua Nowak (ZestStream) with AI coding agents. Not affiliated with or funded by Jeffrey Emanuel; Joshua uses his public tools and is a paying subscriber to jeffreys-skills.md.

Source on GitHub MIT licensed