AgenticWorkflows
Multi-step, tool-using agents: planning loops, tool definitions, guardrails on what an agent is allowed to do autonomously.
Why we need this / what value this brings
Lets an LLM take multi-step actions (call a tool, check the result, decide the next step) instead of just answering in one shot.
When to use this
Only once a single LLM call genuinely can't finish the task — e.g. it needs to look something up, call an API, then use that result.
How to use or implement this
Define a small, explicit set of tools (functions) the model can call, validate every tool call's arguments before executing, and cap the number of loop iterations so it can't run away.
Research questions
- What's the smallest agent loop that's still useful (single tool, single retry) vs a full planner?
- How do you bound an agent's blast radius — read-only tools vs tools that mutate state?
- MCP (Model Context Protocol) as a way to standardize tool definitions across agents.
- Managed agent hosting (Amazon Bedrock AgentCore, running on EC2 under the hood) vs a self-built agent loop on your own compute — see Compute/VMs and Networking/VPC-Subnetting for the networking side of the managed-EC2 case.
Empty folder — drop notes, links, and findings here as you research.