Agentic Workflows

Expert

Design prompt-driven loops for autonomous multi-agent systems.

01 The Concept

Moving beyond single-turn completion, agentic workflows use prompts to orchestrate loops (reflection, tool use, planning) and coordinate multiple specialized agents (e.g. planner, coder, validator).

Each agent is driven by a targeted system prompt and communicates via structured messages, enabling the team to solve complex software engineering or analysis tasks autonomously.

02 Weak vs. Strong

EX 01Critic Agent System Instruction

You are a critical, security-obsessed senior code reviewer.

Your task is to identify security vulnerabilities, logical flaws, and edge cases in the draft code.

Rules:

Be adversarial and nitpicky.
Do NOT praise the author or write 'looks good overall'.
Point out every potential issue, regardless of severity.
Format output: [Line Number] - [Vulnerability] - [Fix].

→ Why it works

Overrides the model's default cooperative bias (sycophancy) to enforce critical feedback, which is key for self-correcting agentic pipelines.

03 Key Points

01Agentic Loops: Prompt-driven loops (plan-act-reflect) that allow models to self-correct and iterate.

02Specialized Prompts: Break tasks down so each agent has a narrow, highly-focused system instruction set.

03Adversarial Roles: Enforce strict critical stances (e.g. nitpicky critic) to avoid polite, unhelpful feedback.

04Structured Handshakes: Standardize agent-to-agent communication protocols (JSON messages or XML feeds).

05State Management: Maintain conversation state and workspace tools accessible by prompts.

04 Model-Specific Notes

Claude is highly effective in agentic loops due to excellent tool use capabilities and strict conformance to XML communication protocols.

05 For Your Role

It's like running a software project where one person writes the requirements, one writes the code, and a third runs the tests.

Next: Prompt Versioning & Git Workflows →