Legacy modernization.
For COOs and engineering leaders. Agentic SDLC.
You are running two programs at once: scale operations to meet demand, and hold together a stack that was built for a smaller company. When throughput stalls and engineering tells you the pipeline is too fragile to change, the decision arrives on your desk. Pause feature delivery and refactor, or wait for coding agents to become good enough to do it for you.
There is now a detailed public answer to that question. Denys Linkov, who led the AI pipeline rebuild at Wisedocs, spent six months consolidating more than ten repositories into a single monorepo, then published the cost-benefit analysis on the AI Engineer podcast. Wisedocs processes medical claims documents that can exceed 10,000 pages, and by 2025 its pipeline was too slow for the customers it had won, too complicated to update, and spread across repositories that, in his words, nobody wanted to touch.
His opening frame is the part worth taking into a board meeting. Technical debt is financial debt. It compounds, and it should only be taken on when the return outpaces the complexity it adds. Linkov put it plainly: introducing additional complexity into a codebase can very quickly outrun the ROI already generated.
Can an agent just do it?
Linkov tested the obvious shortcut. He handed a frontier model the goal, the repository names, the underlying models and the components, and asked it to perform the refactor. It reported completion in 10 minutes and 22 seconds, having written roughly 2,000 lines. Inspection found scaffolding: no implemented models, no deployment path, and a note the model had left in its own output saying it had not added the bootstrap command yet.
"We're still not there where models can self-validate and just one-shot these kinds of problems, but we're getting close."
Linkov expects substantial refactors to become reliably agent-completable within about six months of mid-2026.
Read that as a schedule, not a reprieve. The cost of waiting is falling. The cost of compounding debt is not. What changes hands in the meantime is the scarce skill: not coding, but specification, and the verification infrastructure that makes a specification safe to delegate.
The Wisedocs refactor, in numbers
| Reported measure |
Result |
| Time to reach parity with a stack built over six years |
6 months |
| Repository consolidation |
10+ repositories to 1 monorepo |
| Estimated time to redo the same task on current models |
1/5 of the original time |
| Time to ship features that previously took months |
Under a week |
Reported by Denys Linkov, August 2026. Figures attributed to Wisedocs are that team's reported results, not lowtouch.ai benchmarks. Krithi capability descriptions are ours.
The blueprint, and where Krithi sits in it
The Wisedocs account resolves into four operational decisions. Krithi, the coding agent behind lowtouch.ai Velocity Pods, is built around the same four.
1. Consolidate the repositories
Wisedocs collapsed ten-plus repositories, accumulated over six years, into one. Linkov is clear that modern models can navigate several repositories well enough on their own. The gain is downstream of that: end-to-end testing, verification, deployment and sandbox setup all get significantly harder across multiple repositories, because every agent run spends time cloning and configuring before it does any work.
In Krithi: requirements extraction reads the whole legacy codebase alongside the customer voice before a line is written, and every downstream stage runs in a sandbox built once against that consolidated view.
2. Treat research output as unverified
Choosing an orchestrator took Wisedocs two months, five candidates, 17 criteria and a team of three building proofs of concept, because documentation was not trustworthy enough to decide on. Linkov estimates current research tooling would compress that by 90 percent, and names the risk that comes with it: AI psychosis, where a polished twenty-page report is accepted as fact and the features it describes turn out not to exist.
In Krithi: requirements carry traceability back to their source, and no sprint plan reaches the coding stage without a named human approving it.
3. Keep the human pull request review
Across the six-month refactor, every pull request went through human review. Local automated checks ran first, but the review itself stayed with people, because it is how the team builds a mental model of code it did not write. This is the countervailing force Linkov flags against waiting: AI-native code, written fast and understood by nobody, already reproduces the condition the refactor was meant to cure.
In Krithi: testing runs automated UI, API and regression checks and then opens the pull request. A human reviewer merges to main. The agent does not merge its own work.
4. Refactor in layers
A full-stack rebuild is not the only option on the table. Linkov notes the delivery system can be refactored in layers, isolating parts of the codebase so the high-value pipeline moves while operational systems stay where they are. For a COO, that is the difference between a program with a blast radius and one with a boundary.
In Krithi: a Velocity Pod is scoped to one layer at a time and ships each sprint into the client's own environment, so the boundary is contractual as well as architectural.
One workflow, five gated stages
Every Velocity Pod runs the same agentic workflow end to end, from legacy code and customer input to production, with a human at every gate.
An orchestration agent runs behind every stage below. It keeps ticket and status data in sync with the client's existing Jira, or our own tracker if preferred.
| Stage |
What happens |
Gate |
| Requirements extraction |
Reads the legacy codebase and customer voice, then drafts requirements with full traceability |
Human review |
| Planning |
Turns approved requirements into a sprint plan and per-sprint tickets |
Plan approval |
| Coding |
Builds each ticket into working code, then merges it to the dev branch |
Repeats each sprint |
| Testing |
Runs automated UI, API and regression checks, then opens the PR to main |
QA sign-off |
| Production |
Merged to main by a human reviewer, live in the client's own environment |
Human merge |
What Krithi is
Krithi is the coding agent that powers lowtouch.ai Velocity Pods. It migrates and modernizes legacy applications, builds and maintains data pipelines, stands up greenfield applications, and produces mockups and prototypes. For a product manager it behaves as a virtual engineering team: you bring the requirement, it brings the sprint.
The commercial terms follow the same logic as the workflow. lowtouch.ai is ISO 27001 and SOC 2 Type 2 certified, and contracts are tied to measurable outcomes rather than licenses or utilization. If the ROI does not land, you do not pay.
The decision in front of you
Wisedocs came out of its refactor shipping in under a week features that had previously taken months, with lower pipeline latency, lower infrastructure cost, and developers asking to work in the new codebase. It met 15 of its 17 requirements. Linkov's own summary was that models will keep improving, but that it is sometimes worth pausing, building a monorepo and going ahead anyway.
Waiting is not free, and it is not passive. The teams that can use the fully automated refactor when it arrives are the ones who spent the interval building sandboxes, verification loops and the habit of writing a specification precise enough to hand over. Krithi is that interval, run as a service.
Book a Velocity Pod walkthrough.
Sources
- Denys Linkov, Wisedocs, "Benchmarking coding agents on new vs legacy codebases", AI Engineer podcast. Reported 8 August 2026. Source of the six-month refactor, the 17-criteria orchestrator evaluation, the "AI psychosis" quote, the 10 minutes 22 seconds zero-shot test, and the 15-of-17 outcome.
- Denys Linkov, full talk audio and transcript, AI Engineer.
- METR task-horizon data, cited in the talk as the basis for planning at the 90 to 99 percent accuracy line rather than 50 percent. metr.org
- lowtouch.ai, certifications and outcome-based contract terms, as published August 2026. lowtouch.ai
Rejith Krishnan is founder and CEO of lowtouch.ai. Krithi runs this workflow on client engagements in regulated environments.
