↑
moomooskycow 2 hours ago
1. Internal agent 2. Product features 3. Customer-facing agent
This could all be served by one agent, but might not be as effective as having one focused agent for each need. It's not hard to build shared context for each agent to use, and it's (generally) easier to configure somewhat domain-specific agents than one superagent.
Regarding hosting your own company agent: what's the impetus for self-hosting? If it's a data privacy thing (i.e. you don't want to send data to Anthropic or other model providers), you're going to be somewhat limited. It will cost time and money to set up infrastructure for hosting your own copy of a model, and even if you use SOTA open source models like Kimi K2.5 or Deepseek V3.2 they simply aren't as capable as the SOTA models from the frontier labs.
If hosting your own company agent just means a long-running Claude Code instance on an internal server somewhere that employees and company applications can send requests to, then that's way easier. You might even want to set up an OpenClaw agent (depending on your risk appetite, they're very capable but can be a bit of a security nightmare without proper configuration and management).
SOTA right now is the Claude Agent SDK. Package it up with some company-specific skills that capture and document company idiosyncrasies, what APIs exist and how to use them, hardcoded scripts the agent can invoke for happy path tasks and operations. Compress an index of the skills into your CLAUDE.md file to increase the rate of skill invocation, and build workflows that explicitly invoke appropriate skills as often as possible.
I'm rambling. To wrap up:
1. Internal agent: a hardened OpenClaw agent that employees can chat with, or a Claude Agent SDK integration into Slack is probably best.
2. Product features: "summarize this with AI" is more of a standard LLM request pipeline than an agentic system. Lots of AI enhanced features are just one or more prompts chained together, you can do quite a bit with that and some cleaned-up well-placed corporate context.
3. Customer-facing agent: Claude Agent SDK is, again, probably the best bet if you can swing it.
Again, if you want to avoid sending data to model providers you're gonna have a good amount of infrastructure work to do. I think the OpenAI Agents SDK is more model agnostic, LangGraph too.