Agentic engineering

Overview

Agentic engineering is the practice of building software systems with the help of autonomous or semi-autonomous agents that can plan, write, and execute code to complete tasks. In the newsletter, the term is used primarily in the context of coding agents and the operational patterns required to use them effectively, including the management of agent behavior and the avoidance of common anti-patterns.

For AI Product Managers, agentic engineering matters because it changes how software is specified, built, reviewed, and governed. Instead of treating AI as a simple feature or chat layer, teams must design workflows where agents take actions in real environments, which raises new questions around trust, scope, observability, human review, and safe delegation. PMs working on developer tools, internal automation, or AI-native products need to understand both the upside of agent autonomy and the controls required to use it responsibly.

Key Developments

• 2026-03-05 — Simon Willison’s guide on anti-patterns highlighted behaviors to avoid in agentic engineering, especially letting agents push unreviewed code to collaborators.
• 2026-03-15 — Simon Willison discussed agentic engineering at the Pragmatic Summit in a fireside chat hosted by Eric Lui from Statsig, signaling growing industry interest in the topic.
• 2026-03-16 — Simon Willison published a definition-oriented piece describing agentic engineering as software development assisted by coding agents that can both write and execute code, helping formalize the term in the LLM tooling landscape.

Relevance to AI PMs

• Define safe autonomy boundaries. PMs need to decide which tasks agents can perform independently, which require approval, and what permissions or sandboxing are needed before agents can act on code, data, or production systems.
• Design human-in-the-loop review flows. Agentic engineering introduces failure modes such as unreviewed changes, overconfident execution, and hidden multi-step actions. PMs should specify checkpoints for code review, testing, rollback, and escalation.
• Measure agent performance beyond output quality. In agentic systems, success depends not just on whether the final answer is good, but whether the agent followed the right process, stayed within scope, used tools correctly, and behaved predictably under uncertainty.

Related

• Simon Willison — A key voice in defining and popularizing the term, especially through practical writing on coding agents and anti-patterns.
• Coding agents — The most direct implementation layer for agentic engineering; these are the systems that write and execute code on behalf of users.
• Pragmatic Summit — An event where agentic engineering was discussed publicly, reflecting the concept’s emergence in practitioner conversations.
• Statsig — Connected through Eric Lui’s hosting of the Pragmatic Summit fireside chat, linking the topic to product and experimentation audiences.
• Anti-patterns — Closely tied to agentic engineering because successful adoption depends heavily on avoiding unsafe or poorly governed agent behaviors.