Agentic Workflows

What Makes Something "Agentic"?

An AI workflow is agentic when the model makes decisions about what to do next rather than following a fixed script. The spectrum:

Level	What the AI Does	Example
Completion	Finishes your sentence	Copilot autocomplete
Chat	Answers your question	ChatGPT
Tool use	Calls tools when needed	Claude with MCP
Agent	Plans, executes, adapts	Claude Code full session
Multi-agent	Multiple agents collaborate	CrewAI, AutoGen

Core Patterns

ReAct (Reason + Act)

The model alternates between reasoning and acting:

Thought: I need to find the bug in the test file
Action: Read tests/test_auth.py
Observation: [file contents]
Thought: The test expects a 200 but the endpoint returns 401 when...
Action: Read src/auth.py
Observation: [file contents]
Thought: Found it — the token validation skips expired check
Action: Edit src/auth.py (fix the bug)

This is what Claude Code does naturally.

Planning

Before acting, create an explicit plan:

1. Read the failing test to understand expected behavior
2. Read the source file to find the bug
3. Write a fix
4. Run the test to verify
5. Commit

Claude Code's Plan mode (/plan) does this.

Reflection

After acting, evaluate the result:

Action: Run tests
Result: 3 tests pass, 1 fails
Reflection: The fix broke test_token_refresh. Need to also update the refresh logic.

Self-review in Claude Code's subagent-driven-development pattern.

Tool Use

The foundation — model outputs structured tool calls, host executes:

{"tool": "read_file", "args": {"path": "src/auth.py"}}

Subagent Dispatch

Spawn isolated agents for parallel or specialized work: - Research agent: search codebase, web, docs - Implementation agent: write code, run tests - Review agent: check spec compliance, code quality

Anti-Patterns

1. Over-planning: Planning for 30 minutes before writing 5 lines of code
2. Tool abuse: Calling tools when you already have the information
3. Context pollution: Dumping all search results into main context instead of using subagents
4. Retry loops: Same action, same inputs, expecting different results
5. Premature multi-agent: Using 3 agents when one tool call would suffice