April 15, 2026
From AI-Native Development to AI-Native Platform
How agentic workload changes development platforms — using GitHub as an example
/ AI-native Platform 2026
How agentic workload changes development platforms — using GitHub as an example
How agentic workload changes development platforms — using GitHub as an example
Load unit + platform
Parts 1-2: GitHub platform case.
Part 3: broken load economics.
Part 4: AI-native platform failure.
Parts 5-6: lessons for orgs.
GitHub is a control plane for people and agents
AI is first-class load
+36M in 2025 — 180M+ developers — 80% tried Copilot in week one.
×2 in <2 years — 4.3M AI repositories — AI repos: 230 per minute.
5.5M issues/month — 986M commits in 2025 — Copilot correlates with 8-15% growth.
Intent becomes platform operations
Actions environment + repo clone.
Code search across the codebase.
Draft PR, commits, and session logs.
Self-review, scans, and dependency checks.
Agent HQ, CLI /fleet, Workflows, Mobile
Agent HQ: Copilot, Claude, Codex.
CLI /fleet launches parallel subagents.
Agentic Workflows live in Actions.
Mobile adds background branch changes.
Actions, code review, usage metrics — three signals of new pressure
AI is first-class load
11.5B minutes · +35% YoY — GitHub Actions — 71M jobs/day; agents run on Actions.
60M reviews · ×10 growth — Copilot Code Review — >1/5 reviews; 71% actionable feedback.
First-class traffic — AI-native telemetry — Copilot PRs, cloud agents, CLI tokens.
Better review means higher platform cost
Quality improvements
Stronger reasoning model: +6% positive feedback.
Agentic architecture: +8.1%; 71% actionable.
Cost of improvements
+16% latency on stronger models.
Each review is a platform workload.
Agentic fan-out: one issue spawns a tree of platform operations
One issue -> platform operations tree
Classic Platform
Human changes code, waits for CI.
Metrics: MAU, API QPS, CI, storage.
AI-Native Platform
Intent -> platform action tree.
Fan-out and operations per intent.
New load meets old architecture
TTL + peak hit DB
Apps produced ×10 read traffic.
Settings TTL: 12h → 2h.
Read/write overwhelmed auth cluster.
Causes: isolation, shedding, monitoring gaps.
Agentic workflows started creating avalanche-like platform-side load
April 1 — /agents/sessions degraded due to backend exhaustion.
April 2 — Code Review raised Cloud Agent jobs; ~40% affected.
April 9 — Increased load delayed coding-agent sessions.
Review, CI, Git, agents now fail together
~40% request failures — github.com — At the peak of the March incident.
~43% failed requests — GitHub API — Simultaneously with github.com.
~21% errors — Copilot — Shared failures across multiple surfaces.
Process-wise — integrate; infrastructure-wise — isolate
Unified process, stronger infrastructure isolation
Process Integration
Agents inside PR, CI, audit trail.
Security checks in unified flow.
Infrastructure Isolation
Failure domains stricter than classic SDLC.
Per-tenant isolation and traffic priority.
Discipline becomes visible
DORA 2024: better docs, quality, review speed.
Cost: lower throughput and stability.
DORA 2025: throughput vs stability.
GitHub shows this dynamic.
The right problem class — the right category of actions
Failure domains
User cache → segmented DB.
Audit capacity, isolate key dependencies.
Monolith → isolated services/data domains.
Azure migration: 12.5% → 50% traffic.
Right direction; pressure grows
Fix substrate while load grows.
Multi-provider and mobile agents.
Runtime SDK and remote CLI sessions.
Short-term cautious; medium-term positive.
What to take away
Bottleneck: platform layer, not authoring.
Agentic fan-out spawns platform operations.
Guardrails are production load.
Platform engineering manages fan-out and machine traffic.
tellmeabout.tech · ai4sdlc-research.space
Core sources
Classic PDLC to AI-native article.
GitHub Blog, GitHub Status and Copilot updates.
GitHub Actions, Code Review, org metrics.
DORA 2024 report.
AI-Native Platform
For more materials on this topic, visit the "Book Cube" channel — all links are collected there
Alexander Polomodov, Technical Director & Fellow, T-Technologies
@Book_Cube