The Agentic Loop: Multi-Step AI Workflows
Portal One's agentic runtime transforms your AI agents from single-turn responders into autonomous problem solvers. When you send a message, your agent can chain multiple tool calls, reason about results, and iterate — all in a single conversation turn.
How It Works
Previously, each agent interaction followed a simple pattern: one LLM call → one tool action → wait for approval → repeat. The new agentic loop changes this fundamentally:
- You send a message to your agent
- The agent makes an LLM call and decides which tools to use
- Built-in tools (terminal, files, web) execute automatically in a sandbox
- The agent sees the results and can make another LLM call
- This loops for up to 25 iterations per message
- MCP tools that need approval pause the loop until you approve
- The agent delivers a final response when the task is complete
Real-Time Streaming
You don't have to wait for the agent to finish. Portal One streams every step of the agentic loop in real-time:
- Text content streams inline as the agent thinks
- Tool invocations appear with expandable details showing inputs and outputs
- Iteration counts track how many steps the agent has taken
Auto-Approved vs. Human-Approved Tools
The agentic loop distinguishes between two types of tool calls:
| Tool Type | Approval | Examples |
|---|---|---|
| Built-in tools | Auto-approved (sandboxed) | Terminal, file read/write, web search, code execution |
| MCP tools | Your existing approval workflow | Sending emails, database writes, API calls |
Built-in tools are safe to auto-approve because they execute in an isolated sandbox — they can't affect your systems. MCP tools that interact with external services still require your approval, preserving human oversight where it matters most.
Iteration Limits
Agents can execute up to 25 iterations per user message. This is enough for complex tasks like writing and debugging code, multi-step research, or file processing — while preventing runaway loops.