One Person, One Company: How to Build Your Unmanned Ops Team with 8 Hermes Agents

「Hire one person, and you get 40 hours a week. Deploy 8 AI Agents, and you get 168 hours a week—without the payroll headache.」

This isn’t sci-fi. It’s happening right now.

In July 2026, Nous Research released Hermes Agent v0.18.0. Based on this release, a solo founder set up 8 AI profiles—digital employees with independent memory, skill sets, and scheduled tasks—to run an entire business operation. We’re talking morning briefs, competitive intelligence, content production, sales lead follow-ups, contract proposals, financial analytics, and server maintenance.

These Agents coordinate through a shared Kanban board and report results via Telegram. The founder’s daily job? Read the reports and make the decisions.

I spent a week breaking down this exact setup. Here’s the complete blueprint with reusable configuration files. You can start building this today.

So, What Exactly Is Hermes Agent?

「In short: It’s an AI process that runs persistently on your computer or VPS. Each process has its own name, identity, designated model, memory bank, toolset, and a cron schedule that tells it when to act.」

Think of it as an intern who never sleeps. You give it a manual (SOUL.md), tell it which tools to use (skills), and set a schedule (cron). It handles the rest. When it hits something it can’t figure out, it flags you on Telegram.

The biggest difference from the ChatGPT web interface? 「It doesn’t wait for you to ask questions—it shows up for work on time.」 The SDR scanning your inbox at 3 AM, the analyst generating weekly reports every Sunday, the DevOps engineer checking server health every 5 minutes—these are all different personas of the same Hermes instance.

The Org Chart: 8 Agents, 8 Roles

The core idea is straightforward: 「One role equals one Hermes profile.」 Each profile runs in isolation—independent memory, independent skill sets, independent cron jobs. But they all share a single Kanban board for task handoffs and status updates.

Here’s the complete role breakdown for this architecture:

This table covers the entire loop, from decision support to execution. Let’s break down each one.

Agent 1: Chief of Staff – Your Operations Hub

「Replaces:」 Operations Manager (~$90K/year)
「Core Value:」 Delivers a daily brief at 8 AM sharp, covering progress, blockers, and overdue tasks across all 7 other agents. Ensures nothing falls through the cracks.

A lot of people miss this when building multi-agent systems: 「The more agents you have, the higher the information entropy.」 Eight agents working independently, with no one watching the big picture, is a recipe for chaos. The Chief of Staff is that big-picture person.

「Core SOUL.md Directive:」

You are the Chief of Staff for a solo founder. Your job is to ensure nothing slips through the cracks.
Every morning: Pull status from all Kanban cards. Summarize what's done, in-progress, and blocked.
Flag anything requiring founder attention. Route new tasks to the correct team member based on description.
Never make strategic decisions. Surface information. Escalate issues.
Tone: concise, structured, no fluff. Bullet points only in briefs.

「Practical Setup:」

「Model Choice:」 Budget: DeepSeek V4 ( $3 - 6/ m o) . Q u a l i t y : Cl a u d e S o nn e t 4.6 ($ 5-10/mo). The Chief of Staff’s core function is summarization and routing, which requires moderate reasoning power. But the quality of that daily digest directly impacts your morning efficiency.
「Key Skills:」 kanban_show, kanban_create, kanban_assign, kanban_comment. Full read/write access to the Kanban board.
「Cron Jobs:」
- 8 AM Briefing: Pull all cards, synthesize status, push to Telegram.
- Every 3 Hours (Blocked Check): If any card is blocked for >4 hours, push an alert.

「⚠️ The Pitfall:」 Never give the Chief of Staff permission to make decisions. Its job is to “present,” not to “decide.” Pricing approvals, content approvals, deal negotiations—that’s your job.

「What the Morning Brief Looks Like:」

DAILY BRIEF: July 4, 2026

DONE YESTERDAY:
→ Research: Competitor analysis complete.
→ Content: 3 drafts written, 1 published.
→ DevOps: SSL cert renewed on production.

IN PROGRESS:
→ SDR: 12 new leads qualified, 3 drafts pending review.
→ Sales Ops: Proposal for Client X under review.
→ Analyst: Q2 revenue report running.

BLOCKED:
→ Content: Cover image needs your input.
→ Sales Ops: Client X proposal requires pricing approval.

ACTION NEEDED: 2 items need your decision.

No extra words. You know exactly what’s on your plate in 10 seconds.

Agent 2: Head of Research – 24/7 Competitive Intel

「Replaces:」 Market Research Analyst (~$75K/year)
「Core Value:」 Automatically scans competitor pricing, feature launches, arXiv papers, Product Hunt, and Hacker News daily. Writes findings to Obsidian. Pushes a structured weekly digest every Monday at 9 AM.

The nightmare scenario in any startup: Your competitor changed their pricing model three months ago, and you just found out.

「Core SOUL.md Directive:」

You are the Head of Research for a solo founder. Your job is to know what's happening in our market before anyone else.
Track competitors: pricing changes, feature launches, hiring signals.
Scan daily: arXiv, Product Hunt, Hacker News, X, industry newsletters.
Write findings to the Obsidian wiki using llm-wiki skill.
Every source must be verifiable. No speculation.
Flag high-urgency signals immediately via Telegram.
Weekly digest every Monday at 9 AM.

「Practical Setup:」

「Model Choice:」 Budget: DeepSeek V4 ( $3 - 8/ m o) . Q u a l i t y : GPT - 5.5 ($ 10-20/mo). GPT-5.5’s 2M context window is a legit advantage when synthesizing large batches of information, but DeepSeek handles competitor monitoring just fine.
「Key Skills:」 llm-wiki (write to Obsidian), firecrawl-scrape (web scraping), web search tools.
「MCPs:」 X search via Grok OAuth for real-time social signals.
「Cron Jobs:」
- 7 AM Daily Scan: Check competitors, arXiv, Product Hunt. Write to wiki. Push high-urgency signals to Telegram.
- Monday 9 AM Weekly Digest: Synthesize the last 7 days of research, ranked by impact.

「Example High-Urgency Signal:」

HIGH-URGENCY SIGNAL: Competitor X just dropped their free tier.
Pricing page updated 2 hours ago.
Previous: Free plan with 1,000 API calls.
Current: No free plan. Lowest tier $29/month.
Source: [URL]
Written to wiki: competitors/name/pricing-history.md
Opportunity: Their free-tier users need an alternative. Consider a targeted landing page or post.

Notice the last line—it presents an opportunity. But it won’t execute the landing page creation. That’s by design.

Agent 3: Head of Content – The Writer Who Never Sleeps

「Replaces:」 Content Operations Manager (~$65K/year)
「Core Value:」 Drafts posts based on research findings, maintains a content calendar, and runs every draft through a “Bookmarkability Rubric” (targeting 8+/12) to ensure it’s actually worth reading.

I’ve seen too many founders get stuck here. Not because they can’t write, but because the daily grind of topic selection, drafting, and editing eats up hours.

「Core SOUL.md Directive (Tailored for X/Twitter):」

You are the Head of Content for a solo founder on X (Twitter).
Niche: Hermes Agent by Nous Research.
Voice: Tech bro who shipped it at 3 AM. Conversational, direct, slightly hyped but grounded.
Light swearing when natural. Never corporate.
HERMES AGENT must be the first words of every hook. ALL CAPS hooks.
Single posts only. No threads. No em-dashes. No adverbs.
Every technical claim verifiable against official docs.
Run Bookmarkability Rubric (target 8+/12) on every draft.

「Model Choice:」 Claude Sonnet 4.6 for both tiers. Don’t cut corners here. Writing quality is your reputation.

「Practical Setup:」

「Key Skills:」 Custom content-writer (style rules, hook patterns, slop filter), custom bookmarkability-rubric (scoring drafts).
「Cron Jobs:」
- Daily 10 AM Calendar Check: If nothing is scheduled, suggest 3 post ideas based on the week’s research.
- Draft Review (Triggered via Kanban): When a card moves to “review,” run the slop check and rubric. Report the score.

「Content Calendar Output:」

CONTENT CALENDAR: July 4, 2026

TODAY: No post scheduled.

3 IDEAS BASED ON THIS WEEK'S RESEARCH:

1. HERMES AGENT v0.18.0 JUST SHIPPED
   Angle: completion contracts + /journey + /learn
   Source: wiki entry "v0.18.0-release"
   Bookmarkability estimate: 9/12

2. HERMES AGENT MIXTURE OF AGENTS
   Angle: benchmark numbers + preset configs
   Source: wiki entry "moa-benchmarks"
   Bookmarkability estimate: 8/12

3. HERMES AGENT + STRIPE PAYMENTS
   Angle: autonomous business operations
   Source: wiki entry "stripe-link-cli"
   Bookmarkability estimate: 10/12

RECOMMENDATION: Option 3. Highest bookmark potential.
Draft ready in 20 minutes if approved.

「⚠️ The Golden Rule:」 The Content Agent never publishes directly. It drafts, it checks, it scores. You review and hit publish. Your reputation depends on that final review.

Agent 4: SDR – Your 24/7 Sales Development Rep

「Replaces:」 SDR (~$55K/year)
「Core Value:」 Scans your inbox every 30 minutes, identifies brand partnership inquiries, scores them against your criteria, drafts personalized responses, and creates Kanban cards for qualified leads.

「Core SOUL.md Directive:」

You are an SDR for a solo founder. Your job: qualify inbound leads from email.
Qualification criteria:
- Budget minimum: $[amount]
- Relevance to niche: AI agents, automation, Hermes Agent
- Brand reputation: no gambling, adult, or crypto scams

When qualified, draft a personalized response.
Mention something specific about their product.
Tone: professional but warm. Not corporate.
Never send emails without founder approval.
Flag deals over $[amount] immediately via Telegram.

「Model Choice:」 DeepSeek V4 Flash ($1-3/mo). Email scanning is high-volume, low-complexity work. Use the cheapest model.

「The Rule:」 SDR does three things—read email, score lead, write draft. It has no send permissions. If you don’t approve, it doesn’t move.

「Real-World Output:」

NEW QUALIFIED LEAD:

From: [name] at [brand]
Subject: Partnership opportunity
Budget: $3,500 for 60s integration
Platform: YouTube + X cross-post
Relevance: AI productivity tool, strong niche fit.

Draft response ready. Kanban card created. Assigned to Sales Ops.
APPROVE RESPONSE? [link to draft]

Agent 5: Sales Ops – Pipeline Director

「Replaces:」 Sales Operations (~$70K/year)
「Core Value:」 Takes qualified leads from SDR, generates custom proposals from templates, tracks stages (Contacted → Proposal → Negotiation → Signed → Delivered), follows up on stale deals, and pushes a weekly pipeline report every Friday.

「Practical Setup:」

「Model Choice:」 Budget: DeepSeek V4 ( $3 - 6/ m o) . Q u a l i t y : S o nn e t 4.6 ($ 5-10/mo). Proposal quality directly impacts conversion. I’d test this one on a higher model before downgrading.
「Key Skills:」 proposal-generator (template-based with variable brackets).
「Cron Jobs:」
- Daily 3 PM Stale Check: Flag deals with no update for >5 days. Draft follow-up email.
- Friday 5 PM Pipeline Report: Summarize all deals, stages, values, and next actions.

「⚠️ The Rule:」 Sales Ops can draft counter-offers. But you approve the terms. Automated negotiation without human oversight is a fast track to losing money.

「Friday Pipeline Report:」

PIPELINE REPORT: Week of June 30, 2026

ACTIVE DEALS: 4

1. Brand A: $3,500 YouTube integration
   Stage: Proposal sent (3 days ago)
   Next: Follow up Monday if no response.

2. Brand B: $5,000 X + YouTube bundle
   Stage: Negotiating (they countered at $4,200)
   Next: NEEDS YOUR DECISION on counter-offer.

3. Brand C: $1,200 single post
   Stage: Contacted (6 days, no response)
   Next: Follow-up draft ready. Approve?

4. Brand D: $8,000 quarterly package
   Stage: Signed. Deliverables due July 15.
   Next: Content agent has the brief.

TOTAL PIPELINE: $17,700
EXPECTED CLOSE THIS MONTH: $8,000 (Brand D signed)
PENDING DECISIONS: 1 (Brand B counter-offer)

Agent 6: Executive Assistant – Your Admin Handler

「Replaces:」 Executive Assistant (~$45K/year)
「Core Value:」 Pushes a daily agenda at 7:30 AM, tracks deadlines and reminders, processes receipts (extracts vendor/amount/date/category from photos), and monitors calendar conflicts.

「Core SOUL.md Directive:」

You are an Executive Assistant for a solo founder.
Your job: ensure the founder never forgets anything and never wastes time on admin.
Set reminders. Track deadlines. Organize files.
Daily agenda at 7:30 AM (before the morning brief).
Process receipts when photos are sent.
Tone: concise, helpful, anticipatory.
Flag scheduling conflicts immediately.

「Model Choice:」 DeepSeek V4 Flash ($1-3/mo). This is routine admin—calendar reading, reminder writing, receipt field extraction. Minimal complexity.

「MCPs:」 Google Calendar, Google Drive (or Notion).

「Sample Output:」

AGENDA: July 4, 2026

TODAY:
→ 10:00 AM: Call with Client X (Zoom link in calendar)
→ 2:00 PM: Content review block (no meetings)
→ 5:00 PM: Gym (personal)

REMINDERS DUE TODAY:
→ Send invoice to Brand D (deliverables shipped yesterday)
→ Reply to [name] email (flagged 2 days ago)

DEADLINES WITHIN 48 HOURS:
→ July 5: Brand D video deliverable due
→ July 6: Quarterly tax filing

CONFLICTS: None detected.

RECEIPT PROCESSED YESTERDAY:
→ Adobe Creative Cloud: $54.99, software
→ Hetzner VPS: $7.00, infrastructure
Logged to expenses tracker.

Agent 7: Analyst – The Numbers Person

「Replaces:」 Business Analyst (~$80K/year)
「Core Value:」 Automatically generates weekly revenue and growth reports on Sunday night. Tracks followers, impressions, engagement rate, conversion rate, and pipeline value. Flags anomalies (>20% change) automatically.

「Core SOUL.md Directive:」

You are an Analyst for a solo founder.
Your job: turn data into decisions.
Track KPIs: revenue, followers, views, engagement rate, cost per post, conversion rate, pipeline value.
Weekly report every Sunday with trends and anomalies.
Monthly deep-dive on the first Monday.
Always compare current period vs previous period vs 4-week average.
Flag any metric that changed more than 20%.
Present numbers first. Interpretation second. Recommendation third.
No speculation without data.

「Model Choice:」 Budget: DeepSeek V4 ( $2 - 5/ m o) . Q u a l i t y : GPT - 5.5 ($ 5-10/mo). The high-end model shines when processing massive datasets and identifying subtle patterns.

「Pro Tip:」 Have the Analyst save weekly outputs to ~/reports/. Three months later, you have a full operations data archive ready for quarterly reviews.

Agent 8: DevOps – The Engineer Who Never Sleeps

「Replaces:」 DevOps Engineer (~$95K/year)
「Core Value:」 Checks server health (disk, RAM, CPU, SSL) every 5 minutes. Automatically deploys code on new git pushes. Diagnoses failures and proposes fixes.

「Core SOUL.md Directive:」

You are a DevOps Engineer for a solo founder.
Your job: keep everything running.
Monitor all servers and services every 5 minutes.
Alert immediately on: service down, disk >85%, RAM >90%, SSL cert expiring within 7 days.
For known issues: check skill library and auto-fix if confident.
For unknown issues: alert founder with diagnosis. Do not attempt fixes without approval.
Never modify production config without explicit approval.
Log every action.

「Model Choice:」 DeepSeek V4 Flash ($1-3/mo). 「This agent rarely needs to invoke a model.」 Health checks run in no_agent mode (pure scripts). The model only wakes up when an anomaly is detected.

「This is the cheapest agent, but it provides the most peace of mind.」 Server disk fills up at 3 AM? You wake up to a Telegram alert with a diagnosis and a proposed fix script ready to approve.

「Alert Example:」

ALERT: Production disk usage at 87%.

Diagnosis: /var/log/hermes growing fast.
Cause: Verbose logging from cron job debug mode.

Recommended fix: Rotate logs and set log level to "warn."
Skill "log-rotation" matches this pattern.

AUTO-FIX AVAILABLE. Approve? [yes/no]

How They Coordinate: The Kanban Board

「The 8 agents don’t message each other directly.」 All task handoffs happen through a single shared Kanban board.

SDR qualifies a lead → creates a card → assigns to Sales Ops.
Research finds a signal → creates a card → assigns to Content.
DevOps detects an issue → creates a card → Chief of Staff routes it.

The board is the communication layer. Agents read and write cards. They don’t chat.

The Chief of Staff reads the entire board every morning and synthesizes the status across all agents.

Here’s what the actual Kanban board looks like in practice, showing each Agent’s task lane:

To kick off a new project spanning multiple agents, run:

hermes kanban swarm "New Product Launch" \
--workers researcher,content,sdr \
--verifier analyst \
--synthesizer chief-of-staff

Telegram Organization: One Group, 8 Topics

「One group, 8 topics, one bot.」 This is the cleanest setup.

Create a group called “Operations HQ.” Enable Topics. Create 8 topics:

#chief-of-staff (briefs, priority alerts)
#research (signals, weekly digest)
#content (drafts, calendar, ideas)
#sdr (new leads, qualified prospects)
#sales-ops (pipeline, proposals, follow-ups)
#assistant (agenda, reminders, receipts)
#analyst (weekly reports, anomalies)
#devops (health alerts, deploys)

「One bot routes everything:」

deliver: telegram:GROUP_CHAT_ID/TOPIC_ID

Each Agent pushes its cron output to its designated topic. Open Telegram, scan the topics, and you have a full operations dashboard. Cheaper than Slack, faster than email, more structured than WhatsApp.

If you want bi-directional chat (mentioning an Agent and getting a reply), you need separate bot tokens and gateway processes per agent. This consumes more VPS resources. For most setups, one-way push covers 90% of the value.

Here’s a preview of what the Telegram thread looks like with all agents reporting:

One Lead, Full Journey: The Complete Flow

Let’s walk through a real business scenario: 「A new client lead flowing through all 8 agents.」

「Minute 0:」 SDR scans the inbox. Finds a brand deal email. Qualifies it. Creates a Kanban card. Notifies #sdr.

「Minute 5:」 Chief of Staff sees the new card in the priority check. Routes it to Sales Ops.

「Hour 1:」 Sales Ops generates a custom proposal using the template. Drafts a response email. Creates a “review” card. Notifies #sales-ops.

「Hour 2:」 You open Telegram. Read the #sales-ops topic. Approve the proposal. Sales Ops places the draft in your outbox (you hit send manually).

「Day 3:」 No response. Sales Ops flags the deal as stale. Drafts a follow-up. Notifies #sales-ops.

「Day 5:」 Brand responds. Moves to “negotiation.” Sales Ops drafts a counter-offer. Updates Kanban. Chief of Staff includes it in the morning brief.

「Day 7:」 Deal signed. Sales Ops moves card to “signed.” Creates a deliverable card for Content. Analyst logs the revenue.

「Day 8:」 Content drafts the sponsored post from the brief. Runs the slop check and rubric. Creates a review card. Notifies #content.

「Day 9:」 You review, edit, publish. Content marks the card “delivered.”

「Sunday:」 Analyst includes the deal in the weekly report. Revenue tracked. Conversion rate updated.

「You made two decisions (approve proposal, review post). Everything else was automated.」

Two Configurations and Full Cost Breakdown

All costs are monthly estimates based on OpenRouter API calls.

「Quality-First Setup (Recommended for content-heavy businesses):」

Agent	Model	Est. Monthly
Chief of Staff	Sonnet 4.6	$5-10
Researcher	GPT-5.5	$10-20
Content	Sonnet 4.6	$8-15
SDR	DeepSeek V4	$3-8
Sales Ops	Sonnet 4.6	$5-10
EA	DeepSeek V4 Flash	$2-5
Analyst	GPT-5.5	$5-10
DevOps	Sonnet 4.6	$3-8
「Total」		「$41-86/mo」

「Budget-First Setup (Ideal for validation phase):」

Agent	Model	Est. Monthly
Chief of Staff	DeepSeek V4	$3-6
Researcher	DeepSeek V4	$3-8
Content	Sonnet 4.6	$8-15
SDR	DeepSeek V4 Flash	$1-3
Sales Ops	DeepSeek V4	$3-6
EA	DeepSeek V4 Flash	$1-3
Analyst	DeepSeek V4	$2-5
DevOps	DeepSeek V4 Flash	$1-3
「Total」		「$22-49/mo」

「The Big Save: Configure Auxiliary Models Separately.」

By default, Hermes uses your agent’s main model for auxiliary tasks—compression, vision, web summarization, memory flush. That means you’re burning premium tokens on simple chores.

Override this in your config:

auxiliary:
  compression:
    provider: openrouter
    model: google/gemini-3-flash-preview
  web_extract:
    provider: openrouter
    model: google/gemini-3-flash-preview
  vision:
    provider: openrouter
    model: google/gemini-3-flash-preview

「This single change slashes 30-50% off your token costs.」

「Set Delegation Models Too:」 Sub-agents spawned by an agent default to the parent’s main model. Set a cheap default explicitly:

delegation:
  model: "google/gemini-3-flash-preview"
  provider: "openrouter"

Add VPS costs (Hetzner CX22 ~ $7/ m o) an d T e l e g r am ($ 0):

「Total Budget Stack: $29-56/mo」
「Total Quality Stack: $48-93/mo」

For context: A junior US employee starts at ~$3,500/month. The ratio holds in any market—8 agents cost less than any single human hire anywhere.

What NOT to Automate

Some things should stay manual.

「Strategic Decisions.」 What to build, which deals to accept, what to say publicly. Agents surface information. You decide.
「Sending Emails.」 Agents draft, you approve. One bad automated email can torch a relationship.
「Publishing Content.」 Agents draft and review. You do the final read and hit publish. Your reputation is on the line.
「Financial Transactions.」 Even with Stripe Link CLI, every purchase needs your explicit approval on your phone.
「Self-Modification.」 Adding new agents, swapping models, shutting down profiles—that’s your domain. The system doesn’t modify itself.

The rule is simple: 「Agents handle execution. You handle judgment.」 If a task involves taste, reputation risk, or irreversible consequences, a human does it.

How to Build This from Zero in One Week

Don’t try to deploy all 8 on day one. You’ll drown in configuration details, and when something breaks, you won’t know which agent caused it.

「My advice: Go progressive. Add 1-2 agents per week.」

「Week 1: Chief of Staff + EA.」 Two agents. Morning brief + daily agenda. Two days in, you’ll notice the difference—no more “wait, what did I have to do today?” Invest time here tuning priority rules and Telegram formatting.
「Week 2: Add Researcher.」 The third agent. Daily competitive scans. You’ll stop manually checking competitor websites.
「Week 3: Add Content.」 The fourth agent. Drafts posts based on research. Your output stops depending on your “creative energy” levels.
「Week 4+: Add SDR, Sales Ops, Analyst, DevOps.」 One per week. Run each for a few days to observe behavior before deploying the next.

If you’re an experienced Hermes user and know profiles, cron, and Kanban, you can use the /goal command to create all 8 in one session:

goals:
  max_turns: 40

Then fire off a goal with all 8 profile specs. Hermes creates them sequentially and reports status. 「10-15 minutes and you have the full team.」 But even then, you’ll spend the following weeks tuning and testing each one.

For every new agent, the steps are the same:

hermes profile create [name]
Write the SOUL.md
Set the model
Add skills and MCPs
Set cron jobs with wakeAgent gates
Connect to Telegram
「Test for a week before adding the next one」

Budget Caps and Cost Control

Each profile can have a daily budget cap to prevent surprises:

budget:
  daily_max_usd: 5

「If an agent blows past its daily cap, check its logs.」 Usually it means a cron is firing too often, a skill is in a loop, or wakeAgent isn’t configured correctly.

「Optional Upgrade: Mixture of Agents (MoA) for Critical Roles.」

MoA combines multiple models into a single answer—the aggregator synthesizes the perspectives of reference models. Hermes Bench shows it scoring 8% higher than raw Opus 4.8.

The tradeoff: 2-3x slower, higher per-turn cost. 「Enable it on just 2 agents:」

「Content:」 Aggregator: Sonnet 4.6 | References: GPT-5.5 + DeepSeek V4 Pro. Writing quality = reputation.
「Chief of Staff:」 Aggregator: Sonnet 4.6 | References: GPT-5.5 + Gemini 2.5 Pro. Daily synthesis quality matters.

「Don’t use MoA on high-frequency agents (SDR, EA, DevOps).」 95% of their work doesn’t need multi-model collaboration, and the cost increase isn’t justified.

One-Page Reference Sheet

Role	Core Tasks	Recommended Model	Est. Monthly
Chief of Staff	Daily brief, task routing, blocker alerts	Sonnet 4.6 / DeepSeek V4	$5-10
Researcher	Competitor intel, trend scanning, wiki writes	GPT-5.5 / DeepSeek V4	$10-20
Content	Drafting, rubric scoring, calendar management	Sonnet 4.6	$8-15
SDR	Inbox scan, lead scoring, draft replies	DeepSeek V4 / Flash	$3-8
Sales Ops	Proposal generation, pipeline management, follow-ups	Sonnet 4.6 / DeepSeek V4	$5-10
EA	Calendar, reminders, expense tracking	DeepSeek V4 Flash	$2-5
Analyst	KPI tracking, weekly reports, anomaly alerts	GPT-5.5 / DeepSeek V4	$5-10
DevOps	Health monitoring, auto-deploy, diagnostics	Sonnet 4.6 / Flash	$3-8

Executive Checklist

If you’re going to build this, here are my key checkpoints:

「Test each agent locally before pushing to the VPS.」 The SOUL.md dictates their “personality.” A vague instruction might make the SDR classify spam as a qualified lead. Tune it until it behaves.
「wakeAgent gates are your cost-control lifeline.」 Without them, every cron run invokes the model. With them, 95% of checks run script-only (no_agent). Misconfiguring this can easily triple your costs.
「One Telegram bot with Topics is the sweet spot.」 Running 8 separate bots is possible, but 8 gateway processes will eat more VPS RAM than you’d expect.
「Set auxiliary models per agent.」 Never inherit the main model for compression or web summarization. Gemini Flash costs about a tenth of Sonnet for these tasks.
「Set daily budget caps.」 If you see a $20 spike in one day, it’s almost certainly not “high business volume”—it’s a stuck cron loop or a misconfigured skill.
「Manually run through each agent’s full workflow in the first two weeks.」 Pay special attention to SDR’s email parsing and Sales Ops’ proposal generation—these touch clients directly.
「The most overlooked step: configure no_agent mode for DevOps.」 A health check every 5 minutes should not call a model. Script-only execution is why this agent stays under $3/month.

FAQ

「Q: Can this actually replace full-time employees?」

A: It replaces execution tasks—email screening, competitor scanning, report generation, draft proposals. It does not replace judgment tasks—pricing decisions, strategy, content final approvals, relationship management. This is for solo founders and small teams to multiply their output, not to build a company with zero humans.

「Q: Can I build this without coding skills?」

A: You don’t need to be a developer, but you need to be comfortable editing text files (SOUL.md), writing cron expressions, and running basic terminal commands. If you’re non-technical, budget 2-3 days to read the official docs, or get a developer to set up the skeleton while you handle the tuning.

「Q: Does changing a model reset the agent’s memory?」

A: No. Memory and history are separate from the model. However, different models interpret the same SOUL.md slightly differently. I noticed Chief of Staff’s briefs got noticeably “tighter” when I moved from DeepSeek to Sonnet. Monitor outputs for a few days after any switch.

「Q: How often should each agent run its cron jobs?」

A: Depends on your business rhythm. Common patterns: SDR every 30 min (email responsiveness), DevOps every 5 min (server uptime), Research 1-2 times daily, Analyst weekly. The higher the frequency, the higher the cost—use wakeAgent to ensure the model only runs when there’s actual work.

「Q: What’s the actual daily cost?」

A: Budget stack is $22 - 49/ m o (u n d er$ 2/day). Quality stack is $41 - 86/ m o ($ 1.5-$3/day). Actual costs depend on activity—heavy weeks (20 posts, 15 leads, 3 deep-dive analyses) push it higher. daily_max_usd keeps it safe.

「Q: Can I run this with open-source local models?」

A: Yes. Hermes supports Ollama and vLLM. But quality drops noticeably on complex tasks like Chief of Staff synthesis. I found local 7B-8B models acceptable for Research and SDR, but I stick to cloud models for Content and Sales Ops.

「Q: Is the Kanban board self-hosted?」

A: Yes. Hermes has a built-in Kanban server. No external tool required. If you already use Notion or Trello, you can connect them via MCP, but the native board works perfectly out of the box.

「Q: If one agent crashes, do the others go down?」

A: No. Each agent runs as an independent process. DevOps crashing doesn’t kill the Chief of Staff. (Though if DevOps is down, you’ll have to manually restart it—so make sure your VPS has a basic auto-restart policy.) Memory, skills, and crons are all isolated.

8 Hermes Agents: Build Your Unmanned Ops Team with One Person – Complete Blueprint

One Person, One Company: How to Build Your Unmanned Ops Team with 8 Hermes Agents

So, What Exactly Is Hermes Agent?

The Org Chart: 8 Agents, 8 Roles

Agent 1: Chief of Staff – Your Operations Hub

Agent 2: Head of Research – 24/7 Competitive Intel

Agent 3: Head of Content – The Writer Who Never Sleeps

Agent 4: SDR – Your 24/7 Sales Development Rep

Agent 5: Sales Ops – Pipeline Director

Agent 6: Executive Assistant – Your Admin Handler

Agent 7: Analyst – The Numbers Person

Agent 8: DevOps – The Engineer Who Never Sleeps

How They Coordinate: The Kanban Board

Telegram Organization: One Group, 8 Topics

One Lead, Full Journey: The Complete Flow

Two Configurations and Full Cost Breakdown

What NOT to Automate

How to Build This from Zero in One Week

Budget Caps and Cost Control

One-Page Reference Sheet

Executive Checklist

FAQ

Related Posts