TL;DR

In a 2024 series of letters in The Batch, Andrew Ng argued that agentic workflows, where a model drafts, checks and revises in loops, can matter more than the next model upgrade. He named four design patterns: reflection, tool use, planning and multi-agent collaboration, and judged the first two more reliable than the last two. For marketing operations, that suggests starting with review loops and tool-connected tasks, and treating fully autonomous, multi-step agents as experiments with human checkpoints.

Andrew Ng's argument in brief

Andrew Ng is one of the best-known educators in machine learning. In March and April 2024 he used his letters in The Batch, the DeepLearning.AI newsletter, to set out what he called agentic design patterns, beginning with "Agentic Design Patterns Part 1" on 20 March 2024.

His starting point is that most people use a language model in "zero-shot" mode: one prompt, one answer, no revision. He compares this to asking someone to write an essay from start to finish in one go, without ever going back. An agentic workflow instead lets the model work iteratively: outline, research, draft, critique and revise, much as a human writer would.

To make the case, he cited results on HumanEval, a coding benchmark. GPT-3.5 answered 48.1% correctly in zero-shot mode and GPT-4 67.0%. Wrapped in an agent loop, GPT-3.5 reached up to 95.1%. His point was that the workflow around a model can deliver a bigger jump than upgrading the model itself.

Agentic workflow

a process in which an AI system takes several steps toward a goal, such as planning, using tools, checking its own output and revising, rather than producing a single response.

The four design patterns

  1. Reflection (Part 2, 27 March 2024). The model reviews its own output and improves it. Ng describes a simple version, where the model is asked to critique and then rewrite its work, and a two-agent version, where one agent generates and a second critiques. He calls reflection relatively quick to implement and reports gains in some of his own applications, while being clear that self-critique only sometimes catches problems.
  2. Tool use (Part 3, 3 April 2024). The model is given functions it can call to search the web, run code, or work with email and calendars. His example: for a compound interest calculation, have the model call a Python interpreter instead of guessing the number.
  3. Planning (Part 4, 10 April 2024). The model decides for itself which sequence of steps to take for a larger task. Ng finds this powerful but harder to predict, and notes that workflows with a fixed structure do not need it.
  4. Multi-agent collaboration (Part 5, 17 April 2024). A task is split among agents playing different roles, which can all be the same model prompted differently. He likens it to a manager dividing a project among specialists, and notes that output quality becomes harder to predict when agents interact freely.

Crucially, Ng ranks the patterns by maturity. He found reflection and tool use reliable enough to improve his applications, and considered planning and multi-agent collaboration less mature.

"a less mature technology" — Andrew Ng on planning, The Batch

Why this matters for marketing operations

Much of marketing operations already looks like a workflow: brief, research, draft, review, localise, publish, measure. That makes it a natural fit for agentic patterns, and also a place where errors compound. A single mistaken product claim, copied into twelve language versions and five channels, is far more expensive than one bad chatbot reply.

Ng's framework helps separate what is ready from what is promising. A reflection loop that checks a draft against your brand guidelines is a modest, controllable addition. A tool-connected step that pulls current prices from your product feed instead of letting the model invent them reduces a known failure mode. A fully autonomous agent that plans a campaign, writes the assets and pushes them live is the kind of planning and multi-agent set-up Ng himself flagged as less predictable.

A playbook: introducing agentic patterns, in order of maturity

  1. Pick one repeatable workflow. Good candidates are product descriptions, weekly social posts, campaign reports or translations. Document the current steps and who signs off.
  2. Add reflection first. After the first draft, run a second pass with an explicit checklist: brand voice, banned claims, required disclaimers, reading level, call to action. Ng's two-agent variant (one writes, one critiques) is easy to set up in most tools.
  3. Connect tools where the model would otherwise guess. Prices, stock, dates, legal wording and performance numbers should come from a system of record: your product feed, CRM, analytics or a fixed snippet library, not the model's memory.
  4. Keep the plan human-defined. For now, write the sequence of steps yourself (a fixed workflow) rather than letting an agent invent it. Ng notes fixed workflows do not need the planning pattern at all.
  5. Pilot multi-agent set-ups in a sandbox. If you try role-based agents (researcher, copywriter, editor), run them on internal or low-risk work, and compare their output with your existing process before anything goes live.
  6. Insert human checkpoints before anything external. Publishing, sending emails, changing ad budgets and replying to customers should stay behind an approval step until you have evidence the workflow is reliable.
  7. Measure the loop, not just the output. Track how often the critique step catches real issues, how often humans still correct the result, and time saved per asset.

Limits and common mistakes

The HumanEval figures are from coding, a domain with a clear right answer and automatic tests. Marketing quality is harder to score, so reflection loops need a concrete checklist or they tend to produce confident but cosmetic edits. The benchmark numbers also date from early 2024 and refer to models that have since been superseded; the durable lesson is the workflow principle, not the specific percentages.

The most common mistake is jumping straight to "autonomous agents" because they make good demos, while skipping the boring, reliable patterns. A second is adding loops without measuring them, which adds cost and latency with no evidence of better results. A third is giving agents access to live systems, such as ad accounts or email lists, before access controls and approval steps are in place.

Key takeaways
  • Ng argues iterative, agentic workflows can improve results more than upgrading the model alone.
  • His four patterns are reflection, tool use, planning and multi-agent collaboration.
  • He rated reflection and tool use as more reliable; planning and multi-agent set-ups as less predictable.
  • Marketing teams should start with review loops and data-connected steps, and keep humans approving anything external.

FAQ

Do we need special software to use agentic patterns?
Not to start. A reflection loop can be as simple as a second prompt that critiques the first draft against a checklist. Tool use and multi-agent set-ups usually need an automation platform or developer support.
Which marketing tasks suit agentic workflows best?
Repeatable, well-defined tasks with clear quality criteria and reliable data sources, such as product descriptions, localisation, reporting and content repurposing. Open-ended strategy work is a poorer fit.
Are the HumanEval numbers still relevant?
They illustrate the principle that workflow design can matter as much as model choice. The specific models have been superseded, so treat the figures as a 2024 illustration rather than a current benchmark.

Sources

Design a marketing workflow that checks itself

JRK can help you pick one repeatable marketing process and add reliable AI review and data steps, with humans approving what goes out.

Book a free growth session →

Want this working for your brand?

Book a free 30-minute growth session. We'll look at your funnel and tell you what we'd do first — honestly.

Book a growth session →