ThoughtFoxResources / Case studies / Ocuco
How Ocuco rebuilt its engineering model around agentic coding, and what it means for the organisations that follow.
“We’re now over 95% agent-written code across a 140-strong engineering team.”
A note on independence: Ben Churchill, a ThoughtFox co-founder, is also Ocuco’s Chief AI Officer.
** 99% OF CODE WRITTEN BY AGENTS **
002 29 0000
§01 · The situation
Ocuco is one of the world’s leading software companies in the optical and eyecare sector: a 140-strong engineering and product team supporting thousands of clients across the UK & Ireland, Netherlands, Nordics, and North America. A mature, mission-critical codebase handling PII and PHI in regulated markets, with data governance at the centre of every architectural decision.
In late 2024, Ocuco made a deliberate decision to transform how software was built. Not because the team was failing to deliver, but because the ambition of the product roadmap demanded a step-change in capacity. The organisation spent 12 months building AI fluency, establishing the foundation that would make the next step possible.
In Q4 2025, Ben Churchill, Ocuco’s Chief AI Officer, working in partnership with ThoughtFox, ran proofs of concept using Claude Code against the mature, complex codebase. They confirmed what the team had anticipated: the complexity was not a barrier. The right context infrastructure was the answer.
“We’re now over 95% agent-written code across a 140-strong engineering team.”
By April 2026 the transformation was complete, delivered four weeks ahead of schedule. 19 teams were restructured into agentic pods; GitHub Copilot was retired as the team moved fully to the agentic model, supported by an AI enablement team and regular Communities of Practice.
§01.1 · Scale of change
AI-accepted lines of code grew from 0.03 million in December 2025 to 2.07 million by May 2026, a 20× increase as the agentic model moved from pilot to production across all 140 engineers.
Dec ’25 and May ’26 are measured. The path between follows the growth curve.
§02 · The transformation
The step-change came from three structural changes that together created an entirely different operating model.
Context layer
Ocuco built Hive: an internal MCP server aggregating 37 other MCP servers. It connects live Jira tickets with full comment history, vectorised archives of historic tickets, developer resource libraries, and documentation tooling. Engineers stopped copying context into prompts. Agents understood not just the code, but the ticket, the history, the institutional reasoning. PII and PHI are handled via redaction; Claude Enterprise & Teams plans ensured no data left the organisation. Beyond context, Hive runs active agents across the SDLC: a dedicated bug-triage agent classifies and routes every incoming bug automatically. Hive operates at a 92% success rate across 14,400 calls per 30-day period.
SDLC redesign
The initial Tiger Team revealed the real insight: when AI writes the code, the bottleneck moves. Planning became the critical investment: rich tickets, Claude in planning mode producing complete specs before any code. QA shifted left to product planning. TDD finally became the path of least resistance. Code review became the primary quality gate.
Enterprise configuration
Claude Code out of the box is fundamentally different from Claude Code configured for enterprise software. Ocuco built 126 custom skills, slash commands and agent behaviours calibrated for the codebase, made available to every engineer. Engineers moved from writing code to directing and judging AI output.
§02.1 · Context infrastructure
Hive aggregates 37 MCP servers into one context layer, and runs active agents across the lifecycle. Engineers stopped manually assembling context; agents gained the ticket, the history, and the institutional reasoning, with full observability across every run.
Upstream context
Internal MCP aggregator
37 MCP servers. 14,400 calls / 30 days at a 92% success rate. A dedicated bug-triage agent has auto-triaged 670+ bugs to date, routing every incoming bug the moment it arrives.
Active agents
The insight
The right question is not “can AI handle our codebase?” It is “have we given AI enough context to navigate it?”
§03 · Inside the Tiger Team
The Tiger Team was selected from what were expected to be the most AI-fluent engineers. The first lesson came quickly: assumed AI fluency and actual readiness to work in a fully agentic model are different things. The team needed a few days of structured training to calibrate, but once they did, adoption was fast and genuine.
What happened next was the signal that mattered. Engineers did not just use the new tools: they started building with them. Within the first phase, team members were writing their own skills, commands, and workflows and sharing them across the business. They had stopped being users of an AI system and started being architects of one.
The Tiger Team shipped everything they planned to. Then they ran a bug blitz.
The people reality
Senior engineers with deep expertise were often the most resistant. Not because they feared being replaced, but because they genuinely underestimated what the technology could do. Their scepticism was professional, not personal. Once they saw the capability, many became the strongest advocates.
More junior engineers found it easier to embrace the new model, having less prior framework to unlearn. But seniority turned out to be a poor predictor either way. The real dividing line was character.
“The question isn’t will my experienced engineers accept this? It’s which of my engineers are genuinely adaptable? The answer often reveals more about your team than the transition itself.”
§04 · The bug blitz
The Tiger Team’s first bug blitz: a proof of concept for what agentic bug resolution looks like at pace.
Before
5 daysAverage bug fix time across the Ocuco organisation, historically.After
20–30 minSame categories of bugs, under the agentic SDLC.That is not an improvement in efficiency. It is a different category of engineering capability.
§04.1 · Scaled org-wide · April 2026
When the methodology was rolled out across all 140 engineers, the results scaled with it. The April 2026 org-wide bug blitz closed 404 priority bugs and produced a 50% decrease in skipped tests.
More significant is what happened at the quarterly level. Q2 2026 is the first quarter in Ocuco’s history where bugs resolved exceeded bugs created. That is not a one-off improvement in efficiency. It is a structural inflection point.
For the first time, Ocuco is reducing its total bug backlog while simultaneously shipping more features than at any point in its history. The agentic model did not trade one for the other: it moved both in the right direction at once.
§05 · Breadth of evidence
The headline numbers capture magnitude. The breadth of the evidence is what makes them defensible. In Q2 2026, the organisation delivered 35.8% above commitment, against a 1.6% baseline in the same period one quarter earlier (Q1 2026). A 34-point like-for-like delta.
Like-for-like delta
The Q2 2026 result, 35.8% above commitment versus 1.6% in the identical sprint slot one quarter earlier, is a clean comparison. Same sprint slot, same teams, a 34 percentage-point improvement. The full-quarter forecast confirms this is a sustained shift, not a spike, and it holds across 17 of 20 teams, not a self-selecting top cohort.
Ocuco plans on a quarterly cadence (internally, Planning Intervals): Q3 2025, Q4 2025, Q1 2026, and Q2 2026, the current quarter. All comparisons above are like-for-like across these quarters.
§05.1 · Quality engineering
The common concern with agentic coding is quality regression: more code, shipped faster, with less human scrutiny. The Ocuco data shows the opposite. The team did not just move faster; they rebuilt the quality infrastructure while doing it.
Manual regression is now run by three testers across the entire organisation, down from one to two testers per team historically. The defect detection percentage (DDP) for 2025 was 79%. UAT tickets per deployment fell 49% and regression bugs fell 36%.
Named agents in active use
In their words
“TestRail (a widely-used test case management platform) couldn’t keep pace with how modern QE teams work. Building Quexa on Claude Code saved us over €60,000 a year in licensing and €200,000 in manual effort, but the bigger shift was structural. AI is now a first-class participant at every stage: Claude reads the story, writes the test cases, builds the plan, and flags gaps before release. Not a bolt-on. Built in.”
§05.2 · Engineering performance · DORA
DORA (DevOps Research and Assessment) metrics are the industry’s standard framework for measuring software delivery performance. Across all three key indicators, Ocuco has moved materially.
| Metric | Before | After | Change |
|---|---|---|---|
| Deployment frequencyDeployments per quarter | 42 | 51 | +21%target 60 |
| Mean time to recoverMTTR after a failed deployment | 29 days | 8 days | −72%recovery time |
| Change failure rateDeployments requiring remediation | 42% | 32% | −10ptsfailure rate |
The MTTR improvement is the standout figure. A 72% reduction in recovery time changes the risk calculus for deployments: teams ship more often because the cost of a failed deployment has fallen dramatically.
§06 · The results
Productivity increased by 50% almost immediately. As teams learned the nuances, that grew to 2–3×. The floor was extraordinary. The ceiling kept rising.Iarfhlaith Kelly · CTO, Ocuco
Roadmap recovery
The 850 extra story points of quarterly capacity were not abstract headroom. They absorbed a major 473 story-point market-fit programme into the existing schedule: same headcount, no reduction in force. Two consecutive PIs have now recorded 100%+ sprint completion.
Customer impact
Complex customer issues that previously took weeks of investigation are now resolved in hours. Automated bug resolution, with 55% of bugs fixed with no human intervention, is now a routine operational capability. The €6.1M figure compounds: same headcount, same cost, more output every quarter.
§07 · For every engineering leader
12 months of Copilot and AI assistance produced incremental gains. The step-change came only when the entire SDLC was rebuilt around the assumption that AI writes the code. These are not the same approach at different intensities: they are categorically different operating models.
When AI takes over implementation, the critical work moves to planning upstream and review downstream. Organisations that do not redesign around this create a faster way to produce poorly-specified, under-reviewed software.
Ocuco’s codebase complexity was real. The argument that it was too complex for AI was not: it was an argument about poor context infrastructure, not about AI capability. Hive was the answer. The right question is not “can AI handle our codebase?” It is “have we given AI enough context to navigate it?”
Claude Code configured for enterprise software, with purpose-built skills, workflows, governance design, data controls, and full run observability, is a fundamentally different capability from the default tool. The configuration and operating model matter as much as the underlying model.
What 99% really means
“99% agent-written code” is not a metric about AI. It is a statement about where engineering cognition goes. Ocuco’s engineers are no longer spending their expertise on implementation detail. They spend it on what code does, how it is tested, how it performs, and whether it meets the standard.
Flow Engineering · Virtual CAIO
Ocuco’s transformation is one of the most complete agentic engineering programmes in the industry. 99% agent-written code across 140 engineers. DORA metrics moving materially across every indicator. A quality engineering function rebuilt from the ground up. A roadmap back on track, with capacity to spare.
ThoughtFox partnered with Ocuco to design and deliver the programme. The methodology, Flow Engineering, is available to engineering organisations ready to make the same move.
A Virtual CAIO mandate that rebuilds the SDLC around agentic coding. Hive-style context infrastructure. DORA-based measurement. Tiger-team activation. Day-90 target: 95% agent-written code.
Whether you have licences, pilots or nothing yet, we’ll help you pick a sensible first step.