# What is agentic engineering? Practices and evidence 2026

URL: https://sdclabs.com/blog/agentic-engineering/

> Agentic engineering: AI coding agents write and run the code, engineers write the spec, supervise, test and decide what ships.

# What is agentic engineering? How engineers ship software with supervised AI agents

[Mykhailo Poliarush](https://sdclabs.com/author/poliarush/) 2026-11-197 min read

Agentic engineering is the practice of building software with AI coding agents that write and run code on their own, while engineers set the goal, specify the work, supervise the agents and verify every change before it ships. The agent does most of the typing; the engineer owns the design, the tests and the decision to merge. It is the professional end of AI-assisted development, the opposite of "vibe coding", where nobody reads the code.

## Where does the term agentic engineering come from?

Two posts from February 2026 define the term. On the first anniversary of his "vibe coding" post, Andrej Karpathy wrote that programming via LLM agents "is increasingly becoming a default workflow for professionals, except with more oversight and scrutiny", and that his "current favorite" name for it was "agentic engineering" ([Karpathy on X](https://x.com/karpathy/status/2019137879310836075) , 4 February 2026). "Agentic", in his words, because "you are not writing the code directly 99% of the time, you are orchestrating agents who do and acting as oversight". "Engineering", because there is "an art & science and expertise to it".

Simon Willison started a public guide, Agentic Engineering Patterns, three weeks later. He defines the practice as "developing software with the assistance of coding agents", where coding agents are tools that both write and execute code, and calls code execution "the defining capability that makes agentic engineering possible" ([Willison](https://simonwillison.net/guides/agentic-engineering-patterns/what-is-agentic-engineering/) ).

One clarification: agentic engineering is about how you build software. Building AI agents into your product is a different job, covered in [what is AI agent development](https://sdclabs.com/blog/what-is-ai-agent-development/) .

## How does agentic engineering differ from AI-assisted coding and classic automation?

The difference is who drives the loop and what stops a bad change.

- **AI-assisted coding** (autocomplete, chat): the engineer writes the code and asks for suggestions. The AI never runs anything. Review happens line by line as you type.
- **Vibe coding** : the AI writes code and the person accepts it without reading it. Willison uses the term in its original sense, "coding where you pay no attention to the code at all" ([Willison](https://simonwillison.net/2026/Feb/23/agentic-engineering-patterns/) ). Fine for a throwaway prototype, not for production.
- **Agentic engineering** : the agent plans, edits files, runs tests and iterates toward a goal. The engineer writes the spec, sets the guardrails, reviews the result and decides what merges. Willison puts it at "the other end of the scale" from vibe coding: engineers "amplifying their existing expertise".
- **Classic automation** (scripts, CI pipelines, RPA): deterministic steps written in advance. Anthropic calls LLM systems built this way workflows, "orchestrated through predefined code paths", while agents direct their own steps and tool use ([Anthropic](https://www.anthropic.com/engineering/building-effective-agents) ). Agentic engineering still relies on classic automation: the test suite and CI are what keep the agent honest.

## How does agentic engineering work in practice?

A typical change runs through five steps:

1. **Specify.** The engineer writes what must be true when the work is done: behavior, constraints, edge cases and the files that matter.
1. **Plan.** The agent reads the code base and proposes a plan. The engineer corrects it before any code is written.
1. **Build in a loop.** The agent edits code, runs the tests, reads failures and tries again. Willison's short definition: "Agents run tools in a loop to achieve a goal".
1. **Verify.** Tests, type checks, linters and security scans give a pass or fail. Then a human reviews the diff.
1. **Ship small.** The change goes out as a small pull request with context and evidence that it works.

To build the agents themselves around this loop, with skills instead of job-title roles, see AI agent architecture: skills and a loop.

## What engineering practices does agentic engineering need?

The practices are not new. What changes is that they become mandatory, because an agent produces code faster than anyone can read it.

- **Specs before prompts.** A vague request gets a plausible answer to the wrong question. Write the acceptance criteria first.
- **Tests first.** Willison recommends red/green test-driven development for agents: write the tests, confirm they fail, then implement, because agents "might write code that doesn't work, or build code that is unnecessary and never gets used, or both" ([Willison](https://simonwillison.net/guides/agentic-engineering-patterns/red-green-tdd/) ).
- **A check the agent can run.** Anthropic's own guidance for Claude Code warns that without one, "looks done" is all the agent has to go on, which "makes you the test suite" ([Anthropic](https://www.anthropic.com/engineering/claude-code-best-practices) ; vendor documentation).
- **Evals for AI features.** When the code you ship calls a model, unit tests are not enough. Keep evaluation sets from real cases and run them on every change (see our [AI testing services](https://sdclabs.com/services/ai-testing/) ).
- **Supervision and small pull requests.** "Don't file pull requests with code you haven't reviewed yourself", and "several small PRs beats one big one" ([Willison](https://simonwillison.net/guides/agentic-engineering-patterns/anti-patterns/) ).
- **Guardrails and least privilege.** The [OWASP Top 10 for Agentic Applications](https://genai.owasp.org/resource/owasp-top-10-for-agentic-applications-for-2026/) lists tool misuse and identity and privilege abuse among the top risks. Give agents scoped credentials, sandboxes and no direct path to production.

Supervision is already the norm. In a Stack Overflow pulse survey of about 1,100 developers and technologists in April 2026, 59% used agents at work, but 63% rarely or never let them run on autopilot and 60% block agents from making unapproved system changes ([Stack Overflow](https://stackoverflow.blog/2026/05/27/agents-on-a-leash-agentic-ai-remains-mostly-monitored-at-work) ).

How developers keep AI agents supervised at work, 2026

**Use AI agents at work**

59%

**Prefer predictable single-agent setups**

68%

**Rarely or never let agents run on autopilot**

63%

**Block agents from unapproved system changes**

60%

Source: [Stack Overflow pulse survey (about 1,100 developers and technologists, April 2026; separate questions), published 2026-05-27](https://stackoverflow.blog/2026/05/27/agents-on-a-leash-agentic-ai-remains-mostly-monitored-at-work)

## Where does agentic engineering work, and where does it fail?

It works best on tasks that are well specified, can be checked automatically and fit inside an agent's reliable range. METR measured that range in 2025: models had almost 100% success on tasks that take a human expert under 4 minutes, but under 10% on tasks over about 4 hours ([METR](https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/) ). The length of task agents can finish with 50% reliability had been doubling about every 7 months for six years ([METR](https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/) ; figures as of March 2025). METR's tasks are self-contained and well specified, closer to what "a new hire or a remote internet contractor" could do than to work that needs deep context ([METR](https://metr.org/time-horizons/) ).

What METR found about the length of tasks AI agents can finish

|   | Finding |
|---|---|
| Tasks that take a human expert under 4 minutes | Almost 100% agent success |
| Tasks that take a human expert over about 4 hours | Under 10% agent success |
| Length of task finished with 50% reliability | Doubled about every 7 months over 6 years |
| Same trend on SWE-bench Verified tasks | Doubled in under 3 months |

Source: [METR, Measuring AI Ability to Complete Long Tasks (March 2025; figures as of publication)](https://metr.org/blog/2025-03-19-measuring-ai-ability-to-complete-long-tasks/)

It fails in three predictable ways:

- **Speed you only feel.** In METR's 2025 trial, 16 experienced open-source developers took 19% longer with early-2025 AI tools, while believing they had been faster ([METR](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) ). The sample is small and METR now marks the results as out of date, but the lesson holds: measure cycle time, not impressions.
- **Insecure code.** In Veracode's tests, AI-generated code introduced security vulnerabilities in 45% of cases ([Veracode](https://www.veracode.com/press-release/ai-generated-code-poses-major-security-risks-in-nearly-half-of-all-development-tasks-veracode-research-reveals/) ; vendor research). Security scanning belongs in the loop.
- **Weak foundations.** Google's DORA research calls AI "an amplifier" that "magnifies the strengths of high-performing organizations and the dysfunctions of struggling ones" ([DORA 2025](https://services.google.com/fh/files/misc/2025_state_of_ai_assisted_software_development.pdf) ). Agents on a code base without tests or CI make more mess, faster.

## What does agentic engineering cost?

There is no standard price, so plan around the factors:

- **Model usage.** Anthropic reports that agents use about 4x more tokens than chat, and multi-agent systems about 15x ([Anthropic](https://www.anthropic.com/engineering/multi-agent-research-system) ). Gartner forecasts that by 2028 AI coding costs will overtake the average developer's salary as token use grows ([Gartner](https://www.gartner.com/en/newsroom/press-releases/2026-06-24-gartner-predicts-ai-coding-costs-will-surpass-average-developer-salary-by-2028-as-token-consumption-surges) ; forecast).
- **Review time.** Senior engineers spend more of their week reading and verifying, less writing.
- **Test and CI investment.** Fast, reliable suites and sandboxes are the price of letting agents run.
- **Security and governance.** Scoped credentials, audit logs and scanning.
- **Training.** Writing specs and reviewing agent output are skills the team has to learn.

## What mistakes do teams make with agentic engineering?

- **Merging what nobody read.** Review is the job, not an optional step.
- **Giving agents production access.** Keep them in sandboxes with scoped permissions.
- **Skipping the spec.** The agent fills gaps with guesses.
- **No tests to run.** Without a pass or fail signal, the agent stops when the code looks finished.
- **Huge pull requests.** Big diffs get rubber-stamped. Split them.
- **Judging success by feel.** Track lead time, defect rates and rework before and after.

## How SDCLabs can help

SDCLabs has built software since 2011 and founded the open-source testing frameworks Codeception and CodeceptJS. We work this way every day: AI agents draft specs, code, tests and reviews, senior engineers review and decide what ships, and automated tests verify every change.

Our [AI engineering services](https://sdclabs.com/services/ai-engineering/) set up the same workflow inside your delivery process, from requirements to pull request. If you want agents to take over business processes rather than code, see [agentic process transformation](https://sdclabs.com/services/agentic-process-transformation/) and our wider [AI agentic engineering](https://sdclabs.com/services/ai-agentic-engineering/) services. Engagements start in 1 to 2 weeks. [Tell us about your code base](https://sdclabs.com/contact/) and we will suggest where agents can help first.

Related reading: [how to build AI agents](https://sdclabs.com/blog/how-to-build-ai-agents/) , [AI across the SDLC in 2026](https://sdclabs.com/blog/ai-sdlc-trends-2026/) and [the state of AI engineering in 2026](https://sdclabs.com/blog/state-of-ai-engineering-2026/) .

## Frequently asked questions

### What is agentic engineering?

Building software with AI coding agents that write and run code in a loop, while engineers write the spec, set guardrails, review the result and decide what ships. Simon Willison defines it as developing software with the assistance of coding agents ([Willison](https://simonwillison.net/guides/agentic-engineering-patterns/what-is-agentic-engineering/) ).

### Is agentic engineering the same as vibe coding?

No. Vibe coding means accepting AI-written code without reading it. Agentic engineering keeps engineers in control: specs, tests, review and small pull requests. Andrej Karpathy described it as agent-driven programming "with more oversight and scrutiny" ([Karpathy](https://x.com/karpathy/status/2019137879310836075) ).

### Is agentic engineering the same as AI agent development?

No. Agentic engineering is a way of building any software with coding agents. AI agent development is building agents into your product or business processes; see [what is AI agent development](https://sdclabs.com/blog/what-is-ai-agent-development/) .

### What practices does agentic engineering need?

Written specs, tests written first, a pass or fail check the agent can run, evals for AI features, human review of small pull requests, and guardrails such as sandboxes and least-privilege credentials.

### Does agentic engineering make teams faster?

It can, on well-specified tasks that tests can check. But in METR's 2025 trial of 16 experienced developers, they took 19% longer with early-2025 AI tools while believing they were faster ([METR](https://metr.org/blog/2025-07-10-early-2025-ai-experienced-os-dev-study/) ; small sample, results now marked out of date), so measure cycle time and defects rather than impressions.

#agentic-engineering#ai-coding-agents#software-engineering#test-driven-development

[Mykhailo Poliarush](https://sdclabs.com/author/poliarush/) Founder and CEOLeads SDCLabs, an AI development company founded in Krakow in 2011.[LinkedIn](https://www.linkedin.com/in/poliarush/) [GitHub](https://github.com/poliarush)

## Related posts

[
### How to build AI agents that still work after the demo

How to build AI agents: start simple, give one model a few well-designed tools and clear instructions, then add memory, guardrails, evals and human approval.

AI agentic engineering · 2026-11-17

](https://sdclabs.com/blog/how-to-build-ai-agents/) [
### What is AI agent development, and should you build or buy?

What is AI agent development? Building software where a model plans steps, calls your tools and acts on results. How it works, build vs buy, costs and mistakes.

AI agentic engineering · 2026-11-17

](https://sdclabs.com/blog/what-is-ai-agent-development/) [
### 29% of cloud spend is wasted, and AI is pushing it up

Cloud infrastructure spend hit $167.7B in Q3 2026, up 55%. Companies estimate 29% of their cloud spend is wasted, and outages keep getting costlier.

AI agentic engineering · 2026-11-03

](https://sdclabs.com/blog/cloud-operations-spend-outages/)

## Get an estimate for your product.

[Get an estimate](https://sdclabs.com/contact/) [Book a call with the CEO](https://calendly.com/sdclabs/60min)
