Agentic Workflows

A build pattern where the human specifies the outcome in natural language and an AI agent figures out the steps, tools, and implementation. Contrasts with traditional workflow automation (deterministic, step-by-step node configuration). Four core shifts define agentic workflows: self-healing (the agent diagnoses and fixes its own failures), natural-language control (specs replace nodes), security-by-default (the same model reviews its own code for vulnerabilities), and instant API/MCP integration (the agent reads documentation so you don't have to).

Key points

• Self-healing: the agent fixes its own errors by reading logs, editing code, and updating instructions
• Natural-language control: specifications and conversational iteration replace manual configuration
• Security review: the agent audits its own code for exposed keys, logging of sensitive data, and vulnerabilities
• Instant API integration: the agent reads docs or calls MCP servers instead of wrestling with headers and auth
• Best fit: non-deterministic tasks like research, content creation, lead-gen, customer support
• Worst fit: highly deterministic scheduled processes where n8n remains simpler
• Common failure modes: vague goals, missing 'done' conditions, context rot in long sessions, hallucinated APIs
• The 10-hour course draws a hard line between agentic building and deterministic deployment: self-healing is strongest while Claude Code is actively supervising the build; production workflows still need predictable code, test data, logging, and safe failure paths ^[src-016]
• Agentic workflows reward builders who already understand APIs, webhooks, data shapes, and automation fundamentals, because they can spot when the agent made a poor architectural choice ^[src-016]
• Anthropic's scientific-computing article adds a lab-grade version of the pattern: CLAUDE.md defines goals and constraints, a progress file preserves memory, Git coordinates recovery and review, and test oracles make progress measurable ^[src-072].
• Long-running workflows should be chosen when the task is well-scoped and verifiable; open-ended discovery still needs closer human judgment ^[src-072].
• OpenAI's Codex discussion adds a general-work version: an agent can manipulate files, search documents, create spreadsheets, build web pages, prepare slide decks, summarize a day, and run recurring checks when connected to the user's tools ^[src-081].
• Sio's prompting advice matches the workflow failure modes: vague goals are weak; precise output shape, success criteria, and relevant context make the agent more likely to know when it is done ^[src-081].
• Slash-goal extends agentic workflows toward days- or weeks-long pursuit of hard objectives, including performance improvement, program rewrites, math, physics, and scientific problems ^[src-081].
• QuantumBlack / McKinsey adds an enterprise-process version: agents generate impact when the workflow itself is redesigned around planning, memory, integration, human adoption, and governance, not when a chatbot is attached to each existing step ^[src-111].
• The report's credit-risk memo, market research, legacy modernization, and service-desk examples show that the best agentic workflow candidates are concrete, repeatable business processes with measurable time, quality, cost, or decision-speed outcomes ^[src-111].

Related entities

• Claude Code

Related concepts

• Wat Framework
• Claude Code N8n Integration
• Agentic AI
• Agentic Build Deploy Boundary
• Long Running Scientific Agents
• Test Oracle Driven Agents
• Agent Progress File Memory
• Everyday Agentic Work
• Codex
• Gen AI Paradox
• Agentic AI Operating Model

Source references

• ^[src-005] Nate Herk cluster — Nate Herk — n8n cluster (18 videos)

– Videos referenced: AO5aW01DKHo, 3GAxd90fEE4, tDGiWn0flK8, ZeJXI2MAhj0

• ^[src-016] Nate Herk — "Build & Sell with Claude Code (10+ Hour Course)" (2026-03-12)
• ^[src-072] Siddharth Mishra-Sharma – "Long-running Claude for scientific computing" (2026-03-23)
• ^[src-081] OpenAI — "Codex for Everyday Work: AI Agents Beyond Coding" (2026-05-14)
• ^[src-111] QuantumBlack / McKinsey – "Seizing the agentic AI advantage" (2025-06)

2026-06-27 Codex adoption update

• OpenAI's Codex evidence makes agentic workflows observable at usage scale: the shift is from asking for advice to delegating multi-step work that can inspect files, use tools, execute commands, and modify artifacts ^[src-170].
• The paper reports that the share of individual Codex users submitting a task estimated to require more than eight hours of experienced human work increased nearly tenfold since the start of 2026 ^[src-170].
• This connects Agentic Workflows to Agentic Work Adoption: mature usage looks more like delegation, monitoring, and review than a single prompt-response exchange ^{[src-169] [src-170]}.

Additional source references

• ^[src-169] OpenAI – "How agents are transforming work" (2026-06-25)
• ^[src-170] Drew Johnston, David Holtz, Alex Martin Richmond, Christopher Ong, Prasanna Tambe, Aaron Chatterji / OpenAI – "The shift to agentic AI: Evidence from Codex" (2026-06-25)