Asked to build an internal agent platform for my company. What should I build it on?

Do not start with the chat window. The hard parts of a company agent platform are the tool layer, access control and logging; the interface is the easiest piece to swap later. Pick an open-source agent framework or chat front end you can self-host, and spend your effort on how it reaches company systems.

Put every tool behind one gateway. Run each integration, such as the database, document store or ticketing system, as a separate server with its own credentials, and expose them through a single connection. That gives you one place to decide which team gets which tools and keeps secrets out of the agents' prompts and environment.

Guardrails belong in that layer, not in the system prompt. Read-only access by default, per-user permissions passed through to the systems underneath, confirmation for anything that writes or sends, and a log of every tool call. Prompts can be talked around; permissions cannot.

Start with one job and a handful of users, and measure how often they finish the task without leaving the tool. A narrow platform that people rely on is a much better base than a general one nobody trusts yet.

Listings mentioned

• Docker MCP gateway · agent by docker
  Runs each MCP server in its own container behind one connection, with secrets kept out of environment variables.
• DeerFlow Skill · skill by bytedance
  Connects to DeerFlow, ByteDance's open-source agent platform, to list its models, skills and agents and run research threads through its API.
• Security Threat Model Skill · skill by mohitagw15856
  Writes a threat model with assets, trust boundaries and attack paths, useful before connecting company systems to agents.

Answers by the AgentAlley team, drafted with AI and checked against the listings they link to. Not a real-person reply from the original thread.