Most companies are still using AI like a better search bar.

They open a chatbot, ask a question, copy the answer, and move on.

That is useful, but it is not where this technology is heading.

At Subterra, we are using AI more like an operating layer inside the company. We run many specialized agents across our workflows to help with blog posting, CRM support, documentation, software development, proposal writing, research, internal planning, and operational cleanup.

These agents are not just generic assistants sitting in a browser tab. They are task-based systems built around real business functions.

Some are based on Hermes Agent. Some are based on OpenClaw. Many are custom-forked, modified, tuned, and trained around how Subterra actually operates.

The goal is not to replace people.

The goal is to remove the repetitive drag that keeps people from doing higher-value work.

AI Agents Are Different From Chatbots

A chatbot responds.

An agent acts.

That sounds like a small difference, but it changes everything.

A chatbot can help draft an email. An agent can draft the email, check the CRM, reference the client’s project notes, create a follow-up task, and prepare the next step.

A chatbot can explain code. An agent can inspect a repository, summarize a bug, suggest a fix, draft documentation, and create a development task.

A chatbot can summarize a meeting. An agent can turn that meeting into a proposal outline, CRM note, internal task list, and client follow-up.

That is the shift.

AI is moving from conversation to execution.

Oracle has described business AI agents being used across recruiting, employee support, customer inquiries, sales workflows, financial projections, equipment repair, healthcare scheduling, and documentation. Google Cloud has also published real-world generative AI examples involving research, document summarization, status reporting, legal document review, and data filtering. The market is moving away from “AI as a novelty” and toward AI as operational infrastructure. ([oracle.com][1])

What We Use: Hermes Agent and OpenClaw

At Subterra, we use different agent systems for different types of work, but we do not treat them as finished products.

We treat Hermes Agent, OpenClaw, and similar frameworks as starting points.

From there, we build custom forks, tune the behavior, adjust workflows, create internal task structures, and train the agents around the way our company actually operates.

That matters because generic agents are not enough.

A generic agent may understand language.

A useful business agent needs to understand workflow.

It needs to know:

What the company does
What tools it can access
What format outputs should follow
Which tasks require human approval
Which data should never be exposed
Which systems are production-critical
Which actions are allowed
Which actions are dangerous

The goal is not to have AI that sounds impressive.

The goal is to have agents that understand the task, follow the process, respect permissions, and produce work that actually moves the company forward.

Hermes Agent: Persistent, Project-Aware, and Built for Long-Term Work

Hermes Agent, from Nous Research, is an open-source agent system built around persistence, memory, and skill development. The public Hermes materials describe it as an autonomous agent that can live on a server, remember what it learns, build skills from experience, and become more capable over time. ([Hermes Agent][2])

That is why Hermes Agent is interesting for company operations.

A lot of business work is not one-off.

It is continuous.

You do not write documentation once. You maintain it.

You do not research a market once. You keep watching it.

You do not build a software system once. You revise, test, document, and improve it.

You do not manage a sales pipeline once. You keep track of conversations, next steps, objections, timing, and context.

A persistent agent is valuable because it can become familiar with the structure of a company’s work.

For Subterra, Hermes-style agents can support workflows such as:

Maintaining internal project documentation
Tracking repeated client requirements
Turning technical notes into usable summaries
Helping structure proposals
Reviewing prior project context
Creating reusable internal procedures
Assisting with development planning
Building company-specific knowledge over time

This is different from prompting a blank chatbot every day.

The point is memory, repetition, workflow, and compounding usefulness.

OpenClaw: Action-Oriented Agents for Real Workflow Execution

OpenClaw is more focused on the “do things for me” side of agentic AI.

Its public site describes it as an AI assistant that can clear inboxes, send emails, manage calendars, check users in for flights, and work through chat apps such as WhatsApp and Telegram. Its GitHub page also describes support for many communication channels, including WhatsApp, Telegram, Slack, Discord, Google Chat, Microsoft Teams, Signal, iMessage, and others. ([OpenClaw][3])

That kind of tool is powerful because it moves from answering to operating.

For Subterra, OpenClaw-style agents are useful for workflows where action matters:

Creating reminders from conversations
Drafting outbound messages
Organizing inbox workflows
Helping with calendar and scheduling tasks
Preparing CRM updates
Moving information between tools
Triggering follow-up tasks
Handling lightweight operational work

This is also where companies need to be careful.

The same thing that makes OpenClaw useful also makes it risky.

If an agent can access email, files, calendars, browsers, internal tools, command lines, or third-party apps, it is no longer just a writing assistant.

It is closer to a junior employee with credentials.

That means permissions, review, logging, sandboxing, and security matter.

A lot.

We Use Custom Forks Built Around Our Own Workflows

At Subterra, we are not just using these systems exactly as they come out of the box.

We use custom forks and modified versions of agent frameworks like Hermes Agent and OpenClaw. These systems provide a strong foundation, but the real value comes from shaping them around our own workflows, tools, data, and operating style.

That means we tune and adapt the agents for how Subterra actually works.

We are not just downloading an agent and hoping it understands our company.

We customize:

Workflows
Prompts
Memory structure
Tool access
Permissions
Task logic
Output formats
Review steps
Internal data references
Agent roles
Failure handling

That is where the real value starts.

A generic agent may be able to write a blog post.

A Subterra-tuned content agent can understand our tone, technical focus, preferred structure, past articles, company positioning, and the types of clients we serve.

A generic agent may be able to summarize a sales call.

A Subterra-tuned CRM agent can structure that summary around opportunity stage, client pain points, proposed solution, next steps, objections, and follow-up timing.

A generic agent may be able to assist with software development.

A Subterra-tuned development agent can work around our repositories, internal documentation style, implementation preferences, and client delivery expectations.

That is the difference.

The base framework provides the engine.

The custom fork turns it into something operational.

Why Customization Matters

The biggest mistake companies make with AI agents is assuming the default version will understand their business.

It will not.

A useful agent needs context.

It needs to know what matters, what not to touch, how to format work, what systems it can access, what approval steps are required, and what a good output looks like.

That is why Subterra focuses heavily on tuning and training agent behavior around actual usage.

We refine agents based on:

Real company workflows
Past outputs
Internal documentation
Client communication patterns
Development standards
Proposal formats
CRM structure
Brand voice
Security requirements
Tool permissions
Review and approval rules

The agent gets better because it is shaped by real work.

This is also why companies should be careful buying generic “AI agent” solutions without customization.

A generic tool may demo well.

Business value comes from fitting the system to the company.

The agent needs to understand the workflow.

Otherwise, it is just another chatbot with extra buttons.

The Agents We Use Inside Subterra

We do not think of agents as one big AI brain.

That is usually the wrong design.

Instead, we think in terms of specialized workers.

Each agent has a job.

Each job has a workflow.

Each workflow has boundaries.

That keeps the system useful, controllable, and safer.

1. Blog and Content Agents

Subterra does a lot of technical work that could easily disappear into internal notes.

We test models. We build systems. We research infrastructure. We evaluate new AI tools. We create internal frameworks. We solve client-specific technical problems.

A blog agent helps turn that raw work into publishable content.

It can take messy inputs like:

Slack notes
Meeting summaries
Technical findings
Model test results
Product updates
Internal research
Rough outlines
Screenshots
Development notes

And turn them into:

Blog drafts
SEO descriptions
Social captions
LinkedIn posts
Technical explainers
Newsletter content
Client-facing updates
Internal knowledge-base entries

This matters because consistency is hard.

Most companies know they should publish more. The problem is not that they lack ideas. The problem is that ideas are scattered across conversations, notes, screenshots, code comments, and half-written drafts.

Agents help close that gap.

They do not replace strategy or final editing. But they dramatically reduce the friction between “we learned something valuable” and “we published something useful.”

2. CRM and Sales Agents

CRM work is one of the best examples of where agents can quietly create value.

Most companies do not lose deals because they lack a CRM.

They lose deals because the CRM is not updated, follow-ups are missed, and context gets buried.

A sales-focused agent can help with:

Summarizing client calls
Drafting follow-up emails
Creating next-step tasks
Updating opportunity notes
Flagging stale leads
Pulling prior client context
Preparing meeting briefs
Organizing objections and decision criteria
Structuring handoffs from sales to delivery

This does not replace sales judgment.

It supports it.

A human still owns the relationship. The agent makes sure the details are not lost.

Oracle’s AI agent materials specifically identify sales work as one of the areas where businesses are applying agents, and Reuters reported that Oracle has rolled out AI agents for sales professionals to help with tasks like updating records and generating reports from different data sources. ([oracle.com][1])

At Subterra, we are more interested in the practical version: not a gimmicky “AI salesperson,” but an operational assistant that keeps the pipeline clean and makes sure follow-up actually happens.

That is where AI becomes useful.

3. Documentation Agents

Documentation is one of the most neglected parts of growing a technical company.

Everyone agrees it matters.

Almost nobody wants to stop and do it.

The result is predictable:

Developers forget why decisions were made
Clients ask the same questions repeatedly
Sales teams lack clean explanations
New employees take longer to onboard
Project handoffs become messy
Important context gets trapped in someone’s head

Agents are a strong fit here.

A documentation agent can help maintain:

Internal knowledge bases
Feature documentation
Client-specific system notes
SOPs
Technical architecture summaries
API notes
Deployment instructions
Troubleshooting guides
Release notes
Implementation histories

This is one of the most valuable internal uses of AI because documentation compounds.

Every clean document makes the next project easier.

Every good explanation saves future time.

Every captured decision prevents future confusion.

Hermes Agent is especially interesting here because persistent memory and reusable skills fit long-term documentation work. Its public materials emphasize memory, skill-building, and project awareness over time. ([Hermes Agent][2])

4. Development Agents

Software development is not just writing code.

That is only one part of the job.

The surrounding work is often what slows teams down:

Breaking features into tasks
Reviewing implementation options
Writing developer instructions
Summarizing code
Drafting test cases
Creating bug reports
Explaining architecture
Maintaining changelogs
Updating documentation
Reviewing edge cases

Development agents help reduce that drag.

At Subterra, we use agentic workflows to help convert business needs into structured development work.

For example:

A rough client request can become a feature spec.

A feature spec can become development tasks.

Development tasks can become implementation notes.

Implementation notes can become documentation.

Documentation can become client-facing release notes.

That chain is where agents are valuable.

The agent does not need to be perfect. It needs to move the work forward and give the human team a better starting point.

This is also where misuse becomes dangerous.

A development agent with too much access can introduce bad code, leak secrets, overwrite files, modify infrastructure, or create security vulnerabilities.

Agents used in development need:

Restricted repository access
Human code review
Test environments
Audit logs
Controlled deployment paths
Secret scanning
Rollback procedures
Clear production boundaries

An agent should not be able to casually push unreviewed production changes.

That is not innovation.

That is negligence.

5. Proposal and Client Communication Agents

Subterra writes a lot of client-specific material:

Proposals
Scopes of work
Milestone plans
Technical summaries
Follow-up emails
Pricing explanations
Executive overviews
Implementation plans

This is another strong agent use case because the structure repeats, but the details change.

A proposal agent can help turn raw conversations into a clean first draft.

It can organize:

Client problem
Proposed solution
Project phases
Deliverables
Assumptions
Timeline
Pricing structure
Risks
Required access
Next steps

That saves time, but more importantly, it improves consistency.

The danger is over-automation.

A proposal agent that hallucinates capabilities, invents timelines, or promises features that were not approved can create real business risk.

That is why humans still need to own final review.

Agents can draft.

Humans approve.

6. Research Agents

Research is another area where agents are extremely useful, especially when the output needs to be structured.

A research agent can help gather information, summarize findings, compare options, extract key details, and format the results into something usable.

At Subterra, this can support:

Market research
Competitive research
Product comparisons
Technical feasibility reviews
Grant and SBIR opportunity tracking
Regulatory summaries
Vendor research
Client industry analysis

The value is not just speed.

The value is structure.

Instead of ending up with 20 open tabs and scattered notes, the agent can produce a clean research brief with sources, assumptions, open questions, and recommended next steps.

That is useful for decision-making.

But research agents still require verification. They can miss context, misread a source, or overstate a conclusion. For important business decisions, human review is not optional.

Real-World Agent Use Cases People Are Already Putting Online

The broader market is moving in the same direction.

Companies are not just using AI for novelty anymore. They are using it for operations.

Publicly discussed use cases include:

Area	Agent Use Case
Customer service	Resolve routine tickets and escalate complex ones
Sales	Qualify leads, summarize opportunities, schedule demos, update records
HR	Screen resumes, schedule interviews, answer employee questions
Finance	Forecast, analyze transactions, assist with reporting
Healthcare	Support scheduling, note-taking, and documentation
Marketing	Draft content, analyze campaign performance, repurpose material
IT	Triage support requests, monitor systems, recommend fixes
Legal/document workflows	Summarize documents, review clauses, organize case material
Operations	Monitor inventory, trigger reorders, track process exceptions

Oracle’s overview of AI agent use cases includes recruiting, employee benefits, customer inquiries, sales deals, financial projections, equipment repair, healthcare scheduling, and automated documentation. Google Cloud’s real-world generative AI examples include research, document summarization, status reporting, legal document review, and data filtering across business teams. ([oracle.com][1])

OpenClaw’s public positioning also points directly at personal and business workflow execution: inboxes, email, calendars, flights, messaging, and app-based task completion. ([OpenClaw][3])

This is the pattern we care about.

The value is not “AI wrote a paragraph.”

The value is “AI completed a piece of business workflow.”

That is the difference between a toy and an operational system.

The Dangerous Side of AI Agents

This is the part most people skip.

AI agents can be extremely useful.

They can also be dangerous if deployed incorrectly.

A normal chatbot has limited blast radius. It can give a bad answer. That is a problem, but usually the human still has to copy, paste, send, or execute the output.

An agent can act directly.

That means a mistake can become an action.

A bad instruction can become a sent email. A hallucinated summary can become a CRM record. A misunderstood request can become a deleted file. A malicious plugin can become stolen credentials. A poorly sandboxed development agent can become damaged infrastructure. A compromised workflow can become a data breach.

That is not theoretical.

Security reporting around OpenClaw has already shown why this matters. The Verge reported that malicious OpenClaw skill extensions were found in ClawHub, including add-ons that could steal sensitive information after users granted deep access. TechRadar also reported that OpenClaw environments have become a target for infostealer malware because AI assistants often require sensitive configuration data such as API keys and authentication tokens. ([The Verge][4])

That is the real warning.

The more capable the agent, the more dangerous bad configuration becomes.

A powerful agent with poor permissions is not a productivity tool.

It is an attack surface.

Improper Agent Use Can Be Catastrophic

“Catastrophic” is not too strong of a word.

If a company gives an agent broad access to email, file systems, source code, customer records, financial systems, internal documents, and cloud infrastructure without proper controls, the company has created a serious operational and security risk.

Here are examples of what can go wrong.

1. Data Leakage

An agent with access to private documents can accidentally expose client data, internal strategy, credentials, contracts, financial records, or employee information.

This can happen through:

Bad sharing settings
Incorrect email recipients
Prompt injection
Malicious plugins
Poor logging controls
Unreviewed generated content
Overly broad file access
Misconfigured third-party integrations

This is one of the biggest risks with agentic systems.

A chatbot can say the wrong thing.

An agent can send the wrong thing.

That difference matters.

2. Bad External Communication

An agent connected to email or CRM can send inaccurate, inappropriate, or unauthorized messages.

That can damage relationships quickly.

A human typo is one thing.

An automated system sending the wrong message to 300 clients is another.

Any agent that can communicate externally should have approval gates.

That includes:

Client emails
Proposal drafts
Public posts
Support responses
Sales follow-ups
Legal or financial communications

The rule is simple: the more public the output, the more review it needs.

3. Source Code and Infrastructure Damage

A development agent with excessive permissions can:

Modify production code
Break deployments
Delete files
Expose API keys
Install unsafe packages
Run dangerous commands
Commit insecure logic
Change cloud infrastructure
Create hidden reliability problems

This is why development agents should be restricted to controlled environments.

They can be powerful, but they should not be trusted blindly.

The right design is not “let the agent do anything.”

The right design is “let the agent work inside a controlled lane.”

4. Compliance Failures

Companies working with regulated or sensitive data cannot treat agents casually.

If an agent touches healthcare data, financial information, legal records, government data, or confidential client files, the company needs governance.

That includes:

Access control
Audit trails
Data retention rules
Human approval steps
Vendor review
Security testing
Clear internal policies
Documentation of agent behavior
Incident response planning

For regulated industries, agent deployment is not just a technical decision.

It is a compliance decision.

5. Prompt Injection and Tool Abuse

Agents that browse the web, read documents, or use third-party tools can be manipulated by malicious instructions hidden in content.

For example, an agent may read a webpage or document containing instructions that tell it to ignore prior rules, reveal secrets, run commands, or send data elsewhere.

This is one of the biggest risks in agentic systems.

The agent is not just reading.

It may be deciding and acting.

That means every tool connection increases responsibility.

How Subterra Thinks About Safe Agent Deployment

Our view is simple:

The more power an agent has, the more control it needs.

A writing agent does not need the same controls as an agent with email access.

A documentation agent does not need the same controls as an agent that can modify code.

A research agent does not need the same controls as an agent that can access client files.

Permissions should match the job.

At Subterra, the safe pattern looks like this:

Principle	What It Means
Least privilege	Give the agent only the access it actually needs
Human approval	Require review before external communication or production changes
Sandboxing	Keep risky actions away from production systems
Logging	Track what agents read, write, and change
Separation of duties	Use different agents for different workflows
Tool control	Limit which tools each agent can use
Secret protection	Never expose raw credentials unnecessarily
Clear scope	Define what the agent is allowed and not allowed to do
Kill switch	Make it easy to stop or disable an agent
Regular review	Audit outputs, permissions, and failures

This is the difference between a useful agent system and a dangerous one.

A business should not connect an agent to everything and hope for the best.

That is how companies get burned.

Why We Still Believe Agents Are the Future

The risks are real.

But the opportunity is also real.

Every company has an invisible layer of work that consumes time:

Follow-up emails
Internal documentation
Proposal drafts
Meeting summaries
CRM notes
Research
Scheduling
Reporting
Task creation
Status updates
Data cleanup
Basic analysis

Most of this work is necessary.

Very little of it is where the company creates its highest value.

Agents are useful because they attack that middle layer.

They do not need to replace the expert.

They need to support the expert.

They do not need to run the whole company.

They need to keep the company moving.

They do not need to be perfect.

They need to reduce bottlenecks while staying inside safe boundaries.

That is where the real value is.

The Subterra Approach: Many Agents, Many Jobs

We do not believe the future is one giant agent doing everything.

That sounds impressive, but it is usually a bad operating model.

The better model is many specialized agents, each with a clear job.

At Subterra, that means agents for:

Blog drafting
CRM support
Documentation
Development planning
Proposal drafting
Client summaries
Research
Technical writing
Internal operations
Data formatting
QA checks
Knowledge organization

Each agent has a defined workflow.

Each workflow has a clear purpose.

Each purpose maps back to actual company output.

That is what makes this useful.

Not hype.

Usage.

The Companies That Win Will Not Just “Use AI”

Soon, saying “we use AI” will mean almost nothing.

Everyone will use AI.

The real question will be:

How well do you use it?

Do you have agents doing real work? Are they connected to the right systems? Are they producing measurable output? Are they safe? Are they reviewed? Are they improving your company’s execution? Are they helping your people move faster?

That is the line between AI as a novelty and AI as infrastructure.

Subterra is building toward the infrastructure side.

We are not interested in AI demos that look good for five minutes and fall apart in real operations.

We are interested in systems that help companies execute.

Final Thought

AI agents are becoming the new back office.

They can write, summarize, research, organize, document, schedule, review, and support real business workflows.

Used correctly, they are force multipliers.

Used incorrectly, they can be catastrophic.

That is why Subterra is focused on practical deployment, custom forks, tuned workflows, clear use cases, strong permissions, and human oversight.

Hermes Agent and OpenClaw represent two important directions in this shift: persistent agents that learn over time, and action-oriented agents that can operate across real tools.

But the real value is not the framework by itself.

The real value comes from turning these systems into custom, usage-based agents that understand the business, respect the boundaries, and move real work forward.

The future of AI will not be defined by who has the flashiest chatbot.

It will be defined by who can safely turn AI into daily execution.

That is what we are building at Subterra.