What is Agentic Workflows

Common Patterns

• Tool use: the model picks a tool and its arguments, for example a vector search with a filter
• Planning: the model breaks a request into sub-tasks before acting
• Reflection: the model checks its own intermediate result and retries when it falls short
• Routing: a first call classifies the request and sends it to the right specialist prompt or tool
• Multi-agent: several agents with separate roles hand work to each other

Agentic Workflows vs Fixed Pipelines

• Use a fixed pipeline when the steps are known and the same for every input; it is cheaper, faster and easier to test
• Use an agentic workflow when the right steps depend on the question, such as multi-hop lookups or exploratory search
• Agentic runs cost more: each iteration is an LLM call plus the tool calls it makes
• Cap iterations and log every step, so a run can be replayed and audited

Agentic Retrieval in Mixpeek

Mixpeek's agent search stage runs an LLM reasoning loop inside a retriever. The LLM calls other retriever stages, such as `feature_search`, `attribute_filter` and `rerank`, as tools, for up to `max_iterations` rounds (default 5, at most 20). Agents outside Mixpeek can call retrievers over the MCP server. Related terms: agentic AI, agentic retrieval and agentic RAG.

Common Pitfalls

• Building an agent for a task a fixed pipeline handles
• No iteration cap, so a confused agent loops and runs up cost
• Tools with vague descriptions, which the model then calls with wrong arguments
• No evaluation set, so a prompt change that breaks one task type goes unnoticed