Agent skills in legacy modernization: adoption is real but early — production wins cluster in bounded, human-approved workflows
Asked:
“detailed report of adoption of agent skills for legacy modernizations”
This report covers 37 public evidence rows on roughly 23 named organizations (plus two anonymized clients) using task-specific, tool-using AI-agent capabilities for application and infrastructure modernization, 4 verified offering descriptions, and enterprise-agent context studies — current to 2 September 2026. Case evidence is vendor-published and AWS-heavy; that reflects who publishes case studies, not market share, and no market-wide adoption percentage can be supported from it.
Agentic modernization adoption map — where skills work today, across the lifecycle
Human control gates span every stagereview & approve suggestionseditable artifactschat-negotiated plansexplainable decisionsgenerated tests as verification
Hover a deployment for detail. orange edge = measured, vendor/customer-reported outcome; plain = scope-only adoption signal. Capabilities are drawn from the verified offering descriptions; deployments from the public case rows.
What the map cannot say alone
The strongest quantified result:
Experian cut 49 sprints across seven .NET applications (687,600 lines) — a 47% productivity gain with 80% automation in code transformation.
aws.amazon.comOnly 13 of 37 evidence rows carry a measured outcome; the rest are scope-only signals (e.g.
Coupang, Java upgrades across 700+ applications;
Audible, JDK 17 migration with test-suite expansion).
aws.amazon.comAdoption clusters around repeatable transformations with explicit source and target states — Java 8→17, .NET Framework→.NET 8 on Linux, COBOL/DB2→Java/Postgres, VMware→AWS — not open-ended autonomous ownership.
aws.amazon.comThe one non-AWS quantified case:
CME Group migrated 2,000 sensitive regulatory reports from on-premises Java/Oracle to BigQuery in 3–4 months with Gemini Code Assist.
cloud.google.comTrust remains the brake: 71% of organizations say they cannot fully trust autonomous AI agents for enterprise use — general enterprise-agent context, not modernization-specific.
capgemini.comMeasured outcomes, side by side — different denominators, so no average is drawn
Each bar shows the vendor/customer-reported percentage on its own metric. Time-based and scale-based results are listed below the bars. Sources link on each row.
Verified offering slice — what a modernization agent skill actually bundles
A representative verified slice of publicly described offerings (chiefly AWS Transform), not an exhaustive vendor catalog.
Context: enterprise agents broadly, not modernization penetration
3% → 25%
IBM Institute for Business Value (2,900 executives surveyed globally): AI-enabled workflows expected to rise from 3% to 25% by end-2025. Barriers: data 49%, trust 46%, skills 42%. newsroom.ibm.com
84% / 29%
Stack Overflow 2025 survey (49,000+ developers, 177 countries): 84% use or plan to use AI coding tools, yet only 29% trust them. This is AI coding-tool adoption, not autonomous modernization agents. axis-intelligence.com
71% / 43%
KPMG International: 71% report measurable efficiency improvements and 43% attribute revenue growth to AI initiatives — general agentic-workflow context, not a legacy-modernization metric. assets.kpmg.com
Operating model and maturity
- Pattern: copilot → constrained agent → supervised workflow. The adoption unit is a skill or workflow, not a general-purpose autonomous agent.
- Level 0 manual · Level 1 assistive generation · Level 2 tool-using task agent · Level 3 supervised multi-step workflow · Level 4 governed portfolio orchestration. Most cited production evidence sits at Levels 2–3.
- What works: repository/context ingestion + dependency graphing + domain rules + code edits + test execution + audit artifacts, with human review and rollback as production requirements.
- Evidence gap: strong on acceleration and effort reduction; weak on long-run defect rates, maintainability, security incidents, total program cost, and post-migration reliability.
- Barriers: data/context quality, trust, scarce modernization skills, legacy test gaps, undocumented business rules, security/compliance, benchmark comparability.
Roadmap, decision and KPIs
- 90-day pilot: inventory and dependency baseline (weeks 1–3); pick one narrow, high-volume migration archetype (weeks 3–5); build reusable skills with explicit inputs, tools and policies; run in shadow mode (weeks 5–9); require generated tests and human approval before merge (weeks 9–13).
- Build / buy / partner: buy where a GA agent matches your archetype (.NET, VMware, mainframe); partner for scarce mainframe skills; build only bespoke skills around proprietary frameworks.
- KPI scorecard: cycle time per application, accepted-change rate, escaped defects, rework hours, cost per migrated line, post-cutover reliability.
- Scale rule: scale by archetype, not by repository count alone; keep approval, tests, explainability and rollback in place at every level.
Appendix — full evidence table, all 37 rows
Duplicate mentions (same organization in separate or localized sources) and anonymized clients are flagged; they are not additional unique customers.
| Organization | Vendor / offering | Workflow & scope | Outcome / scale | Type | Source |
|---|
Method: live web research current to 2026-09-02, prioritizing official vendor and customer pages and major research organizations. 37 case-evidence rows (~23 named organizations plus two anonymized clients; duplicates flagged), 4 verified offering rows, 3 analyst-context rows, 5 survey-context rows. Outcomes are vendor/customer-reported, use different denominators, and are never averaged; they may conflate agent contribution with broader engineering and cloud changes. The sample is AWS-heavy because of who publishes case studies; no market-wide modernization-agent adoption percentage is claimed. Long descriptions trimmed for space in the table.