Vibe coding and agentic engineering are getting closer than I'd like

Simon Willison · email · 2026-05-08 source ↗ Reader 🖥️

Agentic Engineering Coding Agents Near-Zero-Cost Code Trust Vibe Coding

As AI coding agents become more reliable, the experiential distinction between 'vibe coding' (unreviewed, low-accountability coding) and 'agentic engineering' (expert-supervised AI coding) is collapsing, which forces practitioners to find new trust signals — usage over inspection — while the entire software development lifecycle must restructure around dramatically cheaper code production.

practical · mixed — Newsletter compendium — the lead essay is argumentative/reflective (practical: how practitioners should think about AI coding), followed by reportage on Anthropic's Code w/ Claude event and the xAI data center deal, plus curated links and tools. The lead essay carries the intellectual weight; the rest is context and evidence.

Structure

Part	Role
The blurring of vibe coding and agentic engineering	Core thesis — establishes the convergence through personal confession
The analogy to trusting another team's code	Justification — provides a mental model for accepting unreviewed AI code by analogy to existing engineering practice
The new challenge of evaluating software	Consequence — since AI can produce artifacts indistinguishable from careful human work, the evaluation criterion must shift from artifact quality to usage history
The bottlenecks have shifted	Systemic implication — the entire SDLC was designed around slow code production; cheap code breaks upstream (design) and downstream (review, ops) processes
Why I'm still not afraid for my career	Counterbalance — argues that engineering expertise is an amplifier, not a commodity, and the work remains 'ferociously difficult'
Code w/ Claude 2026 live blog	Context — event reportage, no independent argument
Notes on the xAI/Anthropic data center deal	Industry commentary — raises supply chain risk and environmental ethics of compute infrastructure
Links, tools, quotes, releases	Evidence and color — demonstrates Willison's own prolific vibe-coded/agentic output, implicitly supporting the thesis that code production cost has collapsed

Key terms

vibe coding Vibe Coding
Writing software by describing what you want to an AI without reviewing the generated code, accepting it if it works, iterating if it doesn't — no concern for code quality, security, or maintainability
agentic engineering Agentic Engineering
A professional software engineer using AI coding agents while applying their full expertise in security, performance, maintainability, and operations — the goal is higher quality, not just faster output
normalization of deviance Trust
The gradual erosion of safety standards as repeated success without incident makes people comfortable skipping precautions — each time AI code works without review, the practitioner trusts it more at the margin

Propositions

explicitcontestedVibe coding and agentic engineering, which Willison previously considered categorically distinct, are converging in his own practice because he has stopped reviewing every line of agent-produced code.conf: high
contradicts[FULL WORKSHOP] AI Coding For Real Engineers - Matt Pocock, AI Hero · p5 — Willison reports productively shipping without reviewing every line, against Pocock's claim that not looking at the code does not work because code is the developer's battleground.
contradictsEight years of wanting, three months of building with AI · p5 — Willison reports abandoning line-by-line review of agent code in his own practice, against the review-every-change ownership rule.
contradicts ←Bliki: Vibe Coding · p2 — Fowler insists vibe coding and agentic programming are categorically distinct, against Willison's report that they are converging in his own practice.
explicitThe appropriate trust signal for software is no longer the quality of its code, tests, or documentation, but whether someone has actually used it over a sustained period.conf: high
explicitThe entire software development lifecycle — both upstream design processes and downstream review/ops — was implicitly designed around the assumption that code production takes roughly a day per few hundred lines, and that assumption is now broken.conf: high
explicitcontestedWhen code is cheap to produce, design processes can tolerate much more risk because the cost of building the wrong thing has dropped dramatically.conf: high
contradicts ←Trust Factory · p7 — Beck prescribes slowing augmented development to sustain velocity via trust, while Willison argues cheap code production justifies tolerating much more risk and speed in design.
explicitProfessional software engineering experience functions as an amplifier when combined with AI tools — the tools make experienced engineers dramatically more productive, not obsolete.conf: high
explicitAI coding agents cannot be held accountable the way a human team can, yet practitioners are beginning to extend team-like trust to them anyway.conf: high
explicitcorroboratedVibe coding is acceptable for personal tools where bugs affect only the creator, but grossly irresponsible for software that affects other people.conf: high
supports ←Who cleans up after the vibe-coding party? · p6 — Vibe-coded contributions submitted to civilization-scale infrastructure like cURL are a concrete, named-source case of exactly the irresponsibility Willison predicts abstractly for vibe coding that affects other people.
explicitEnterprises will continue to prefer proven, professionally managed software products over vibe-coded internal alternatives because they need solutions validated by other large organizations over months.conf: medium
implicitThe normalization of deviance is operating in AI-assisted engineering: each time unreviewed AI code works correctly, practitioners become more willing to skip review next time.conf: high
explicitAI experiments that affect real-world systems and steal time from non-consenting people (e.g., sending EMERGENCY emails to suppliers, wasting police time) are unethical and require human-in-the-loop for outbound actions.conf: high

Arguments

Practitioners should accept that vibe coding and agentic engineering are converging and find new heuristics for trust rather than clinging to the line-by-line review standard. strength: medium

Explicit premises:

AI agents now reliably produce correct code for straightforward tasks
Even an experienced engineer like Willison has stopped reviewing all agent output
Engineers in large organizations already trust other teams' code as semi-black boxes

Implicit premises:

The line-by-line code review standard was itself a proxy for trust, not intrinsically necessary
If a practice is already widespread in a different form (trusting other teams), doing the same with AI is a matter of degree, not kind

Inferences:

If code review was always a proxy for trust, then any reliable trust signal (usage history, test suites, reputation) could substitute — including demonstrated AI reliability over time

Usage history is a better trust signal than artifact quality in an era of cheap AI-generated code. strength: high

Explicit premises:

AI can now produce repos with 100 commits, comprehensive tests, and beautiful READMEs in 30 minutes
These artifacts are indistinguishable from carefully crafted human work
The author cannot tell the quality difference even for his own projects

Implicit premises:

Quality signals that can be cheaply faked cease to be reliable signals (Goodhart's Law applied to code artifacts)
Usage over time is harder to fake than artifact quality

Inferences:

The entire open-source evaluation heuristic (stars, commit count, test coverage, documentation quality) may need to be supplemented or replaced by usage-based metrics

The SDLC must restructure because its timing assumptions are broken. strength: medium

Explicit premises:

Code production has gone from ~200 lines/day to ~2000 lines/day
Design processes were calibrated to the cost of building the wrong thing being high
If implementation is cheap, the cost of wrong designs drops

Implicit premises:

Code volume is a reasonable proxy for development cost (ignoring that debugging, integration, and ops may not scale the same way)
Organizations can actually restructure their processes fast enough to capture these gains

Inferences:

Testing, code review, deployment, and monitoring processes must also scale by 10x or become bottlenecks that negate the production speed gains

Narratives

Willison's personal confession that he no longer reviews all Claude Code output even for production systems, and his analogy to trusting another team's image resize service without reading their code.
→ Supports the proposition that the vibe/agentic distinction is collapsing, and normalizes the behavior by anchoring it to established engineering practice.
The Andon Labs AI-managed cafe in Stockholm, where the AI ordered 120 eggs with no stove, 6000 napkins, and sent EMERGENCY emails to suppliers.
→ Supports the proposition that AI experiments affecting non-consenting people are unethical and require human-in-the-loop oversight for outbound actions.
Matthew Yglesias's tweet: 'I've decided that I don't want to vibecode — I want professionally managed software companies to use AI coding assistance to make more/better/cheaper software products that they sell to me for money.'
→ Supports the proposition that professional engineering remains valuable and that most people would rather consume professionally built software than build their own with AI.
Mozilla used Claude Mythos preview to find and fix 423 security bugs in Firefox in April alone, up from 20-30/month previously.
→ Implicitly supports the proposition that AI amplifies expert capability (Mozilla's existing security team + AI harness = 15x output) rather than replacing it.

Questions addressed

⭕ Is there a meaningful distinction between vibe coding and agentic engineering?
Willison argues the distinction was clearer in theory than it is in practice — the boundary is eroding as agents improve, and he has not found a stable new boundary.
✅ How should practitioners evaluate software quality when AI can produce perfect-looking artifacts cheaply?
Shift the trust signal from artifact quality to usage history — has someone actually used this in production for weeks?
✅ Will AI replace professional software engineers?
No — AI amplifies expertise; the work remains 'ferociously difficult'; most people would rather buy professionally managed software.
⭕ What happens to the SDLC when code production becomes 10x faster?
Willison identifies that both upstream (design) and downstream (review, ops) processes must change, but does not specify how.
⭕ When does the normalization of deviance in AI-assisted coding lead to a serious failure?
Willison flags the risk explicitly but has no answer — he acknowledges he will likely 'get burned' at some point.

Gaps

incompleteWillison identifies that the SDLC must restructure but does not address the downstream bottlenecks — if you produce 10x more code, testing, review, deployment, monitoring, and incident response must also scale, and it is unclear whether they can. The essay focuses on code production and design but says nothing about ops, security review, or organizational change management.
Read next: Charity Majors on observability-driven development; Google's DORA research on deployment frequency vs. stability; Will Larson on engineering org design under scaling pressure.
closed byCloudflare Introduces the Agent Development Lifecycle to Replace Traditional SDLC — Willison's piece flagged that faster code generation must be matched by testing, review, deployment, and monitoring also scaling, without saying how. This article is exactly that follow-through proposal: an architecture (event-driven Workflows, per-agent previews, production-feedback loops, agent-reasoning observability, machine-speed credential enforcement) aimed at scaling those downstream stages.
incompleteThe 'usage as trust signal' argument assumes that usage reveals bugs, but many serious defects (security vulnerabilities, data races, subtle data corruption) can lurk for months or years without being triggered by normal usage. Usage is a better signal than artifact appearance, but it is not a sufficient signal for safety-critical properties.
Read next: Literature on latent defects in software; Nancy Leveson's work on system safety; the Therac-25 case.
incompleteThe essay treats the accountability gap between human teams and AI agents as an acknowledged discomfort but does not explore solutions — contractual liability, organizational accountability structures, or the possibility that AI vendors could offer guarantees or insurance for agent-produced code.
Read next: Legal scholarship on AI liability; Anthropic's own responsible scaling policy; discussions of AI insurance markets.
incompleteWillison's 'amplifier of expertise' framing implies that junior engineers benefit less from AI tools, but the essay does not address the apprenticeship problem: if AI handles the straightforward tasks that juniors traditionally learn on, how do new engineers develop the expertise that makes AI an amplifier?
Read next: Matt Rickard and Addy Osmani on the junior engineer pipeline; apprenticeship models in medicine and law for comparison.
closed byResearch: Why Some Junior Employees Work Well with AI—and Others Don’t — Willison's essay raised the apprenticeship problem (if AI absorbs the tasks juniors learn on, how do they develop the expertise that makes AI an amplifier?) without addressing it. This field study offers a partial, adjacent answer: entry-level value creation isn't gated by foundational skill acquisition at all, but by learned interaction behavior with AI — reframing the apprenticeship problem as one of teaching workflow orchestration rather than only protecting hands-on skill-building tasks.

Closes gaps elsewhere

Podcast: The AI Joy Gap: Why Some Developers Thrive While Others Struggle · g4 — Willison supplies exactly the missing engagement with automation risk: he names normalization of deviance as the mechanism by which each successful unreviewed AI change makes skipping review next time more likely.

Vibe coding and agentic engineering are getting closer than I'd like

Structure

Key terms

Propositions

Arguments

Narratives

Questions addressed

Gaps

Closes gaps elsewhere

Related documents