SUPPORT HOMEWORK ·Support Engineer, Nous Research
ⓘ Independent job-application page. Not affiliated with, endorsed by, or operated by Nous Research. Public data as of 30 September 2026.

Homework for my application to Nous Research.

How I would run support for Hermes Agent and the Nous Portal, measured on the live queue, and a working support console built on public data.

2,302
new issues in 7 days
17%
of sampled issues with a team reply
26 / 43
debug-share links already gone
141k
Discord members
00

Summary

The queue

About 2,300 new issues a week and 14,187 open. Bots label and prioritise. In a 196-issue sample, 17% got a team reply and 60 of 90 closes came with no comment.

The job

Be the human loop on top of the bots: first responses, evidence that survives, clean escalations, and the paid surfaces (Portal, billing, Business).

The work sample

A console that reads the queue live, triages any issue in seconds and writes the reply and the escalation, plus a Portal API error doctor.

TakeawayWhy it matters for support
Response coverage is the gapLabelling is automated. What moves user trust is a person replying within a day, starting with P0 and P1.
Evidence expiresDebug-share links are temporary; capture what matters into the issue on first read.
API errors should name the causeA missing key currently reads as a bad model name. Small error-text fixes remove whole ticket classes.
Escalate once, completelyOne packet with version, repro, surviving evidence and impact count replaces fifty duplicate pings.
01

Nous Research in 2026

An open-source agent at very large scale, with a paid inference and tools layer growing on top. Each surface brings its own tickets.

SurfaceWhat is publicSupport read
Hermes AgentMIT-licensed agent, launched Feb 2026: 250k stars, 53k forks, releases several times a month. CLI, TUI, desktop (macOS, Windows), 10+ messaging gateways, many providers and backends.Install, update, gateway and desktop issues dominate.
Nous PortalOpenAI-compatible inference at inference-api.nousresearch.com/v1 plus a tool gateway. Plans Free, Plus $20, Super $100, Ultra $200 a month with credits.Auth, credits, model availability, upstream incidents.
Hermes Business and EnterpriseManaged and self-hosted team offerings (SSO, admin, spend), first rollouts under way.SLA-shaped tickets arrive with the first customers.
CommunityDiscord with 141k members; GitHub issues (Discussions off); docs site with an FAQ.Public questions, scams, incident chatter.
ResearchHermes models, Psyche decentralised training on Solana, Atropos, DisTrO.Crypto and token questions need a fixed script.
02

The queue, measured

GitHub API, 30 September 2026. The demo recomputes these live.

MeasureValueRead
Open issues14,187Plus about 32,900 open PRs.
New issues, last 7 days2,3022,120 closed in the same week: close to flat.
Issues 2 to 7 days old with any comment52%Half of new reports sit silent for days.
Open with no comment4,571The backlog of unanswered reporters.
Team reply (sample of 196, 2 to 15 days old)17%9% within 24 h; median 19 h when it happens. 83 reporters heard only from other users.
Closed with no team comment60 of 90Fixes land; the reporter is not told what changed.
Closes by one account69%Triage is concentrated; a support owner spreads it.
Opened vs closed, 8 weeksnet +336 to +1,408 a weekBacklog grew in 7 of 8 weeks; 14 to 20 Sep closed 443 more than it opened.
needs-repro open967Where evidence completeness pays off first.
Open P0 / P111 / 48The daily must-touch list.
Largest open areasgateway 2,627 · desktop 1,835 · install-update 976 · auth 806Where playbooks and docs fixes go first.
Leaked secrets in 300 recent issues0Debug-share redaction works; keep scanning anyway.
03

Competitive map, support angle

Open agents compete on community trust as much as features. How peers carry their queues.

ProjectOpen issuesModelSupport angle
Hermes Agent14,187Open agent + paid Portal and Business.Benchmark. Highest volume in the set.
OpenClawabout 9,000Largest open agent community (391k stars).Same scale problem, no paid inference layer.
Clineabout 1,500IDE-embedded coding agent with an enterprise tier.Discussions on; narrower surface.
OpenHandsabout 900Coding agent with a hosted cloud.Smaller backlog, community chat.
Gooseabout 400Foundation-run desktop and CLI agent.Discussions + Discord; no bundled inference.
OpenRoutern/aModel router.Upstream for many Portal-style questions.

Issue counts from the GitHub API, 30 Sep 2026.

★

Findings from building the console

Each creates a steady ticket class; each has a small fix.

FindingEvidenceSuggested fix
Missing key reads as a bad model nameA model from the live /v1/models list, called without an Authorization header, returns 400 "Unknown model" (3 of 3 tested). With a bad key it returns 401.Return 401 for a missing header. Until then, one FAQ line.
Debug reports expire before triageThe bug template requires hermes debug share links, which are temporary by design. In issues filed 20 to 29 Sep, 26 of 43 links were gone; 19 of 28 issues had none left.Copy the key lines into the issue at first read (the demo flags expired links), or keep pastes until the issue gets a first response.
Hermes models missing from the Portal listNo Hermes or Nous model id in the public list of about 430 models; nousresearch/hermes-4-405b answers "retired".A retirement note with the recommended replacement in the docs and the error text.
No public status pageNone found for the Portal or the tool gateway.A status page cuts "is it down?" tickets during upstream incidents.

Checked 30 September 2026 against the GitHub API, the live Portal API and hermes_cli/debug.py.

04

The JD duties, and my plan for each

JD dutyHow I would do itDetailed in
Own requests from first response to resolutionDaily pass: P0/P1 without a reply, then issues 2 to 7 days old. Every thread gets an owner and a next-update time.§05, demo 1
Investigate, reproduce bugsReproduction environments for the top surfaces (Windows, WSL2, macOS desktop, Telegram and Discord gateways, Portal auth).demo 2
Implement straightforward fixes and workaroundsSmall PRs for docs, error text and config defaults; workarounds documented on the tracking issue.★
Partner with engineers on larger fixesOne escalation packet per problem, with surviving evidence and an impact count from duplicate search.§06, demo 2
Build tools and automationExtend the existing sweeper bots with evidence checks, duplicate suggestions and first-response drafts; humans post.§07
Establish metrics and goalsHuman-response coverage, evidence completeness, needs-repro age, duplicate and reopen rates, Portal ticket mix.§05
Recurring issues into product, docs and processWeekly trend note: top clusters, counts, one proposed fix each.§05
Shape a growing support organisationWritten playbooks, escalation rubric and on-call handoffs that a second hire can pick up in a week.§08
05

Metrics and first goals

MetricDefinitionFirst goal
Human first responseNew issues with a maintainer or support comment within 24 h, bots excluded.P0/P1: all within 4 working hours. Everything: 80% within 48 h.
Evidence completenessBug reports with version, OS, repro and a readable debug report.Capture evidence at first read.
needs-repro ageMedian age of open needs-repro issues.Under 14 days.
Duplicate rateNew issues closed as duplicate.Top clusters get a pinned issue or a docs fix.
Reopen rateClosed issues reopened within 14 days.Below 5%.
Portal ticket mixAuth, billing, model availability, rate limits, upstream.Weekly note with one fix per category.
06

Escalation standard

LevelWhenAction
P0Secret exposure, data loss, security, or a core flow down for most users.Page engineering now; public status update.
P1Core flow broken for a platform or provider, no workaround.Same-day engineering owner.
P2Broken with a workaround.Document the workaround; schedule the fix.
P3/P4Minor or cosmetic.Batch weekly.

An escalation is complete when it has the exact version, environment, minimal repro, failing output, evidence that will still be there tomorrow, an impact count and what support already tried.

07

Automation in the queue

WorkflowHowGuardrail
Evidence checkOn open: template gaps, version, debug-link liveness, secrets.Deterministic; comments only what is missing.
Duplicate suggestionSearch by exception text and title terms; link candidates.A person confirms before closing.
First-response draftModel rewrites from the checks' findings only.Human posts; never asks for secrets.
Trend noteWeekly clusters by label and exception.Counts from the API; prose reviewed.
08

First 30 / 60 / 90 days

DAYS 1 TO 30 · ANSWER
  • Daily P0/P1 pass; reproduction setups for the top surfaces.
  • Baseline the metrics in §05.
  • Discord support channel rules and scam script.
DAYS 31 TO 60 · TOOL UP
  • Evidence checks and duplicate suggestions on top of the bots.
  • Escalation template in use by everyone.
  • First weekly trend note.
DAYS 61 TO 90 · SCALE
  • Status page proposal; error-text fixes from ★.
  • Portal billing playbook with the billing owner.
  • Playbooks a second hire can run.
09

Method & sources

Built from the public job posting (2026), the Hermes Agent repository and GitHub API, the live Portal API and model list, the Hermes docs and public press, all read on 30 September 2026. The demo is my own tool and reads public data only. Unsolicited interview homework; happy to walk through any section.

Repo: NousResearch/hermes-agent · issue templates

Debug share: hermes_cli/debug.py

Portal: /v1/models · docs

Docs: FAQ

Company: TechCrunch, Jul 2026 · careers

Demo: nous-support.leverlabs.workers.dev

Independent support homework for the Nous Research Support Engineer role · 2026 · edwardtay.com