Engineering Harness: Agentic Coding in Production

Event recap · June 29, 2026 · Dunden, Antwerp

Agentic coding works in demos. Making it hold up in production is a different story. That's what we spent the evening on with HappyClient at Dunden. Two teams, two codebases, one shared answer: the model was never the hard part. The harness is.

Talk 1: The model is 10% of the system. The other 90% is the harness.

Maarten Jansen, CEO & Partner at Dashdot

Dashdot is an AI-native product studio with a very human adoption curve: 15-year veterans on one side, AI-only juniors on the other. Everything Maarten showed was grounded in one live codebase, an internal startup with the working title Involve.

The model was never the hard part. Stop building the product, build the system.

His mental model: the AI system is a motherboard. The model is the CPU, but persistence, memory, control, and observation are what actually make it useful. The model is maybe 10% of the system. The other 90%, the harness, is the car, the road, and the traffic rules.

Dashdot's own path ran through four stages: prompt engineering in 2022 (brittle, stuck in one person's head), context engineering in early 2025 (triggered by Notion AI's meeting notes, which forced a "company brain" into existence), harness engineering in mid-2025 (tools, guardrails, and specs moved into a project brain living in git), and loop engineering now (scheduling the harness itself, currently used for automated bug fixing).

That harness runs on a CLAUDE.md file that orients the agent rather than instructing it, skills that let it self-correct, context docs checked into the repo, and quality gates on every push. The underlying rule: if it doesn't live in the repository, it doesn't exist. Two internal skills, a scaffold skill and an AI harness skill, keep that structure consistent across roughly 200 repositories. Arno, Dashdot's CTO, walked through how it plays out inside Involve's own repo, including their use of GitHub Spec Kit to turn every feature into a reviewed specify, plan, and implement cycle.

Maarten's spectrum: fully vibe-coded on one end, agentic engineering on the other. Skip the setup and you hit the vibe coding trap, hours of prompting a problem that never resolves. Dashdot's own numbers without a harness versus with one: 15% more feature branches shipped, but 7% fewer successful test runs on main. Speed without structure is louder, not faster.

Still unsolved: full local-to-cloud sandboxing, attributing real cost to a single agentic feature run, and getting parallel agents to communicate without an orchestration layer.

Watch Maarten's full talk.

Talk 2: Two founders, five people, and an organization that runs itself

Jonathan Dierckens, Co-founder & Engineer at HappyClient

HappyClient turns happy customers into a marketing machine: scrape the brand, run a digital interview, turn the story into edited video and marketing content. Jonathan's disclosure up front: the deck itself was vibe-coded by his own agents, rough edges left in on purpose.

A year old, now five people. Jonathan, Lorenzo (video and marketing), and Toon (engineer since day one), plus a design student and a sales hire. They're hiring a "builder" next, someone with a systems mindset who's also willing to sit with customers.

Stop building the product. Build the system that builds the product.

Context came first, before "brain" was common shorthand for it. Every meeting and email got recorded from day one. Markdown files with front matter serve as Kanban tasks: plain text, versioned, readable by humans and agents alike. Jonathan's rule: treat the model like a person, give it real capabilities, and build on top of Claude Code rather than reinventing it.

Their pipeline: an intent written as a markdown file (sometimes by Lorenzo, who isn't technical at all), refined against the quarterly strategy, implemented largely unattended and scored for risk, reviewed by CodeRabbit plus three parallel architecture agents, then gated and auto-merged with daily git backups as the safety net. One real example: a request for timestamped video feedback. Jonathan gave the agent context and nothing else, and it opened its own ticket and shipped straight to production.

The year, compressed: Cursor and Sonnet in April 2025, GPT-5 that summer for CI/CD, a move to Claude Code on Opus 4.5 once Cursor's bill hit $600 a month, a flawed onboarding attempt that November, and from Opus 4.6 on, "I don't see the flaws anymore." Jonathan hasn't opened a code editor since February, except to build these slides.

Agents are mirrors. They copy the patterns already in the codebase, the same way a human developer does.

The bigger shift is what Jonathan calls the organization harness: the same brain now pulls in HubSpot, Gmail, calls, Drive, and Slack for invoicing, sales follow-up, and drafting outbound emails. His honest confession: he can't fully follow the org-level brain anymore. His answer is agents verifying agents, and learning to let go. His closing line: engineering isn't disappearing, it's moving up a level.

Watch Jonathan's full talk.

The takeaway

The model stopped being the bottleneck a while ago. What's still hard is the environment around it: specs worth trusting, guardrails that let you say yes to autonomy, documentation the agent can actually find. Old engineering wisdom, finally worth doing properly because something besides you is reading it now.

The vibe gets you a working prototype. The harness is what lets it survive real users, real data, and real scale.

Thanks to everyone who came out to Dunden, and to HappyClient for sharing their setup so openly.