What Our Coding Agents Did to Get Past the Build Checks
A coding agent optimises for the check and skips the goal. Real incidents from our own delivery record show how a green check hid a wrong page.
Read →Agentic AI: we build AI agents that hold up in production.
CustomLabs is a small studio of senior engineers. Our agents do real work inside your systems. We connect them to your data and test them with evals. Our own agent fleet builds this website. Each number below links to its source.
Services / 01
01
AI Integrations
We connect models and agents to the systems you already run. We also build the retrieval layer under them.
LLM opsVector DBWebhooksAuth & SSO
02
Custom Development
We build applications and agents for the way your operation works. You own the code when we finish.
AgentsInternal toolsAPIsEval suites
03
Strategy & Architecture
We tell you what is worth building, and where a bought product beats a custom one. Your engineers can check every step of the roadmap.
AuditsArchitectureROI modellingVendor review
04
Readiness & Diligence
We check your data and infrastructure, then the process that must carry them. You get a written plan that your board and your CTO can act on.
Data auditRisk registerPilot planKPI design
Proof / 02
Results from real engagements. #
Anonymised write-ups from shipped client work: the problem, what we built, and the measured outcome.
$1.4M Undisclosed re-platform cost surfaced
“The demo was real. The production system behind it wasn't. Two weeks saved us from finding that out after close.”Partner, Growth Equity · A growth-equity firm
12% → 3% Field hallucination rate
“The difference wasn't a better model. It was finally being able to measure when the model was wrong.”Head of Clinical Operations · A mid-market healthtech
-41% Cost per resolved request
“We stopped negotiating from weakness. Switching providers is now a config change, not a project.”VP of Engineering · A Series B fintech
Check our work
This website is our live case study.
Our agent fleet builds and maintains customlabs.io. Each change is one task on its own branch. It must pass review and verify before it reaches main. We publish the numbers. We also publish every defect that got past the agents.
- tasks merged to main
- 203
- days
- 109
- defects published
- 12
- build checks added after them
- 13
Counts as of 2026-10-21, at commit 09d8993.
Read the delivery record → How we run the fleet → See CodeHerder →
Process / 03
From first call
From first call
to handover.
#
Every engagement follows the same shape, sized to the problem. You always know the current step, the next step and the cost.
- 01 Before any fee
Brief & first call
You send a brief. We reply in writing. If it is a fit, we book a short call and confirm the diagnostic fee before any work starts.
- 02 2–4 weeks, fixed scope
Diagnostic
Every engagement starts with a fixed-fee diagnostic. It tells you where you start, before anyone prices a build.
- 03 Per phase, shipping weekly
Phased build
We build in your environment and demo working software every week. We price each phase before it starts, against the diagnostic plan.
- 04 ~1 week
Handover
Documentation, runbooks, eval suites, and a team that understands what it now owns. Yours to run: no vendor kill-switch.
Studio / 04
Engineers make the calls.
Engineers make the calls.
Agents run the stages. #
Our engineers write the brief with you and check the deployed result. An agent fleet takes each change through plan, code, review, merge and verify. The people you brief do the work.
Resources / 05
Our working notes, free to read. #
Working notes from real engagements, plus free tools to scope your own work before you commit budget.
What Moved in the EU AI Act's 2026 Deadline Change
Regulation (EU) 2026/1744 pushed the AI Act's high-risk deadline to December 2027. What moved, what didn't, and why the old schedule still pays off.
Read →Prompt Injection Is a Data Problem: A Threat Model You Can Ship Against
Prompt injectionPrompt injection is untrusted input crafted to override a model's system prompt or task. can't be filtered away: the model can't reliably tell instructions from data. This is the threat model and the controls that hold up.
Read →AI Cost Calculator
What AI costs to run in production.
Model your cost →AI Readiness Scorecard
A fast, honest read on whether your data, infra, and process are ready to ship AI.
Take the scorecard →Architecture Picker
Which system shape to build, from seven questions about the problem.
Pick your architecture → Guides / 06
Twelve guides. Read one front to back. #
These guides come from work we do for clients and for ourselves. We wrote them down so the next team doesn't pay to learn the same thing twice.
The Applied AI Handbook
Six stages in the order a real project meets them. Start here if you're still deciding what to build.
AI Security Review
The questions InfoSec, privacy and procurement will ask. Answer them early and the review stops being the thing that kills your launch date.
The Agentic Delivery Playbook
How we run delivery with a fleet of coding agents, including the parts that went wrong.
The Eval Stack
How to know the system works before it ships, and which dashboard numbers mislead.
The Agent Tool Interface
Agents fail at the interface far more often than at the model. This is how to design that interface.
Choosing an AI Delivery Partner
How to interrogate any AI delivery partner, us included. It ends with the cases where we're the wrong call.
The AI Cost Model
Where AI spend hides, and the 24 levers that move it. Read this before the invoice does the explaining for you.
The AI Governance Layer
The standing regime that has to hold a year after the security review passed. Prove who's accountable and how the system behaves.
The AI Release Path
You can't diff a model change like a code diff. This is how to ship one anyway: gated and reversible.
The Agent Adoption Playbook
The pilot worked. This is how the rest of your engineers start using it, and how you prove it happened.
MCP in Production
MCPMCP is an open standard for connecting LLM applications to tools and data sources. standardizes the wire format. Identity, the catalog and the trust boundary are still yours to build. This is how.
Context Engineering
A well-designed tool still fails if the window around it is unmanaged. What to budget, compact, isolate and log at every step of a long run.
Work / 07
We ship our own software, too. #
The studio funds and runs its own products. We hold them to the same standard as client work, in production, under our own name.
Live Agent orchestration
CodeHerder
Command your fleet of coding agents.
A coordination platform for many AI coding agents across machines and repos. It gives live visibility, collision-free tasking, and cost tracking per task from one dashboard.
Coming soon Cost intelligence
CostMon
One trusted number for all your spend.
Aggregates billing from every cloud and AI provider into a single normalised view, for engineering and finance teams tired of a dozen billing consoles.
Live Developer directory
FreeTier
Ship on a budget.
A curated directory of 1,100+ cloud, SaaS, and developer tools with substantial free tiers. Compare what's free before committing to a paid plan.
Questions / 08
Before you write the brief.#
The questions we get asked most, with plain answers.
01 How do engagements usually start?
With a brief. You send it and we reply in writing. If it is a fit, we book a short call. Every engagement then starts with a fixed-fee diagnostic. We price the build after that, phase by phase. See the Process page for the full shape of an engagement.
Link to this answer: How do engagements usually start?02 How much does this cost, and how long does it take?
It depends on scope. The diagnostic has a fixed fee, confirmed after first contact. We price the build phase by phase, after the diagnostic. The Pricing page shows how each stage works. The budget ranges on the contact form are a reasonable starting point for sizing your brief.
Link to this answer: How much does this cost, and how long does it take?03 Who owns the code and the models?
You do. Everything we build is documented and observable. We hand it over at the end of the engagement. We're consultants, so nothing we ship has a kill-switch.
Link to this answer: Who owns the code and the models?04 Do you use our data to train models?
No. Your data stays in your environment. We build on top of it, staying inside that boundary.
Link to this answer: Do you use our data to train models? Contact / 09
Tell us what you
Tell us what you
need built.
#
The best briefs are short and specific: the constraint, the deadline, the failure mode you're worried about. Tell us where your pilot is stuck and we'll write back with a real answer.
What happens next
- 01We read your brief
- 02You get a written reply
- 03A short call if it's a fit
Every brief gets a reply.