2026 (rolling) — Agentic engineering: the umbrella

Why this is its own entry

“Agentic engineering” is the umbrella discipline under which nearly every other entry in this repo lives. Distinct from “prompt engineering” (which 2026 has retired as the load-bearing term — see the CLAUDE.md + plan mode entry) and broader than “harness engineering” (which is one part of agentic engineering), it names the practice of designing, building, and operating AI agents as production systems.

The term crystallised across 2026 alongside the conference circuit (see the AI Engineer Europe entry), and the agentic-engineering open-standard skills repo positions itself explicitly under this banner. Senior practitioners use it as the umbrella category their work falls under; recruiters use it to describe the role; conference tracks organise around it.

What it actually covers

A working definition: agentic engineering is the practice of engineering reliable behaviour from non-deterministic AI agents. The core skills cluster into seven sub-disciplines, every one of which has at least one dedicated entry in this notes set:

Sub-discipline Covered in
Harness engineering — the layer around the LLM (context, tools, memory, control loop, gates) harness engineering entry
Context engineering — what goes into the model’s context window, when, and how it’s recycled CLAUDE.md + plan mode entry — references the Anthropic Effective context engineering post
Skill engineering — packaging reusable instructions / capabilities for cross-context activation skills + marketplace entry
Hook / control-flow engineering — deterministic enforcement of guarantees the LLM alone can’t provide hooks + effort entry
Multi-agent orchestration — when to dispatch, isolate, reconcile across agents managed agents entry, Hermes Agent entry
Spec-driven workflow design — three-document (spec → plan → code) discipline with human review between each CLAUDE.md + plan mode entry
Evaluation and feedback — how you tell whether the agent is doing the right thing without re-reading every output covered in passing across the entries; no dedicated entry yet

This taxonomy isn’t authoritative — different practitioners group differently. But the shape of “there are sub-skills, and they cluster” is converged: the Latent Space crowd, the Anthropic engineering blog, the Cognition Labs writeups, and the ThursdAI / AIE Europe speakers all use roughly these categories.

Why this matters for the rest of the repo

Two practical implications:

1. Skills compound

If you’re spending time on Claude Code, you’re not just learning “a tool.” You’re acquiring the agentic engineering skill set, which transfers across tools. The CLAUDE.md discipline ports to Cursor rules; the spec-driven workflow ports to Codex CLI; the hook patterns port to any tool that exposes a settings.json-like extension surface. The harness-engineering principles port to Hermes Agent, OpenCode, anything ACP-compatible (see the Hermes entry).

This is consistent with Tom Blomfield’s economics framing (see the Blomfield entry): the leverage you get from agentic-engineering skill compounds across tools, which is why the per-user payoff is larger than what previous dev-productivity skills returned.

2. The discipline is what survives model churn

Models change every few months in 2026. The specific Anthropic / OpenAI / DeepSeek release isn’t durable. The discipline of agentic engineering is. A CLAUDE.md you write today will work against GPT-5, Gemini 3, or the next DeepSeek release — because the skill of writing standing instructions for a non-deterministic agent doesn’t care which agent. The spec-driven workflow doesn’t depend on which CLI you’re using. The hook-pattern discipline doesn’t care which harness exposes the hooks.

This is why the notes-set is dated by month: implementations churn; the discipline doesn’t. Six months from now, the specific Anthropic-changelog entries will be different. The practice of structured context loading, spec-before-code, hook-based enforcement, and isolated subagent dispatch will be recognisable.

The career-shape signal

A worth-flagging meta-observation from the AIE Europe debrief + the Blomfield observations + Anthropic’s own internal data:

“Agentic engineer” is solidifying as a job title in 2026. Job listings, recruiter outreach, conference speaker bios all use the term. The shape: someone who treats the LLM as a component to be engineered around, not a service to be prompted.

This isn’t a new role — senior software engineers using AI agents for production work were doing this in 2024 — but the name matters for hiring discoverability, for compensation benchmarking, and for telling a coherent story about what skill set you’re investing in.

What we’re not covering (yet)

Evaluation infrastructure is the obvious gap in this repo right now. The agentic-engineering discipline has a fast-moving sub-field around how do you tell whether your agent is doing the right thing (LLM-as-judge, structured evals, behavioural red-teaming, regression tests across model versions). This deserves its own dedicated entry once the patterns stabilise.

Try this

The exercise that makes the taxonomy concrete: pick a single problem you’ve solved with Claude Code in the last week. Walk through each of the seven sub-disciplines above and ask “how did this manifest in that work?”

The exercise isn’t “score yourself.” It’s make the implicit explicit. The discipline is doing the seven things deliberately; the failure modes are doing them implicitly and getting random results.

Sources