I think startups are a great fit. Getting a really good agent out of the box lets you scale and give your customers value fast. All you need to think about is the business logic: system prompts, tools to give the agent, skills, etc. You won't need to spend time on building the infra layer, orchestration loops, memory, implementing automations, etc.

Developers with customer-facing chat products are the ideal customer.

If a startup has a specific flow they want the agent to take and their traffic is bursty, then I'd recommend using a framework like Mastra and deploying onto a sandbox.

For long-running always on agents where it's important to learn the users preferences overtime, our approach is the highest ROI.

Good question. There a few differences between our approach and shipping an agent with the Claude agent sdk.

1. Our approach has cron-based or trigger-based automations built-in. Building automations with claude agent sdk requires setting up separate infrastructure.

2. Our approach has self-learning built-in. Building a feature like "dreaming" https://docs.openclaw.ai/concepts/dreaming with claude agent sdk also requires setting up separate infrastructure.

3. Our approach decouples the harness and the compute, which lets developers enforce a stricter security boundary, while claude agent sdk ships with the harness, shell, and filesystem in one process https://platform.claude.com/cookbook/claude-agent-sdk-07-hos....

4. Our approach does not vendor lock developers.

You could pick the latest harness and then switch when another better one rolls out. Our bet is that a developer's time is better spent speaking to their customers than switching harnesses.

> what are the cost and security implications?

Cost is the token usage and container uptime.

> One Docker container per-customer sounds like it would be really expensive.

The advantage is per-user memory and self-learning. For context, Claude Managed Agents uses one sandbox per session: https://platform.claude.com/docs/en/managed-agents/environme....

> Are they started on-demand, or run 24/7?

24/7 (best for customer-facing chat products).

> What keeps users from using the agents for general purpose tasks, protects against prompt-injection, etc?

Users define their agent with a system prompt, tool definitions, and skills (which separate a media generation agent from a people search agent). We use Openrouter which has a prompt injection detection feature: https://openrouter.ai/docs/guides/features/guardrails/prompt....

The most valuable pieces of information an AI agent startup can gather is access to their customer's proprietary data and knowledge of their customers preferences (memory + self-learning).

Even as the cost of writing code goes to zero, those two pieces of information are non-commodities.

Thanks for the feedback. The main idea is that today to built a best-in-class agent, developers build the agent loop, session management, tools, memory, skills, automations (cron + trigger-based), sandboxed deployment, and self-learning.

By providing Hermes with a system prompt, custom tools, and skills, developers get the agent loop, session management, automations, sandboxed deployment, and self-learning for free.