AI и продуктивность в Яндексе: что реально заработало / Андрей Попов

Summary (business/strategy focus)

Andrey Popov (Yandex) describes how Yandex scaled internal AI/agent tools from early external prototypes into an organization-wide “new standard” where agents are used daily across engineering and beyond. The company’s current focus is increasing agent adoption and shifting engineering from manual coding to agent-driven execution, while simultaneously evolving infrastructure and operating models to handle higher load and more autonomous workflows.

---

Implementation roadmap & organizational shift (timeline)

• Early stage (1H 2025): External AI tools appear → employees prototype and code with them; internal research/tools emerge; early “agent” building blocks form.
• Second stage (2H 2025): It becomes clear that existing instruments alone aren’t sufficient → Yandex invests in improving quality; “agent use” becomes a teachable skill.
• Breakthrough (start of 2026 / “this year”): Agents become effective at coding and agent scenarios; universal agents catch up to narrow tools, making specialized tools less necessary.
• Current & next: Adoption is already “settled,” so Yandex now shifts toward more end-to-end automation and agent-mediated operations, expecting major movement toward more autonomous stages soon.

---

Core operating model: agents as the default workflow

• Adoption is the main KPI (described as already exceeding benchmarks).
• Training & enablement:
  • 50 MCP servers providing access to internal infrastructure (quality not yet top-tier; usage rising).
  • Internal “ambassadors” who consult teams on implementation.
  • Certification/skills profiling for employees (engineers) to benchmark capability and assign near-term goals.
• Agents used for “all tasks,” not just coding:
  • Automate routines; support processes like on-call duty and experiment evaluation.
  • Use cases span: new projects, product improvements, legacy redesigns, migrations, and operational/incident work.

---

Key metrics / KPIs mentioned (with targets & interpretation)

• Agent adoption

  • Metric: (number of people using agents) / (people with working days).
  • Result: >100% excluding vacations → implies agents are used even on vacations/personal automation.

• Time-saving

  • Approach: compare “user + agent” vs “no AI” and multiply time boost per activation by number of tool uses.
  • Result: about ~8% pure time savings attributed to tool usage.

• Productivity proxy (PR output)

  • Result: AI users produce ~50% more pull requests than non-AI users.
  • Limitation noted: non-AI users are nearly gone, so comparisons have diminishing validity.

• Overall productivity lift (research findings)

  • Company trend: 3–5× (or more) productivity increase for intensive “active agent users.”

• Coding automation share

  • ~85% of code lines generated by agents.
  • Concentration goal concept (“75^3 developers”):
    • Target framing: 75% of developers generate 75% of pull requests
    • Current state: ~60% (60% engineers effectively stopped manual coding).

• Infrastructure/workload growth expectations

  • Infrastructure goals: 1) Prepare for at least 10× workload increase from agents 2) Shift infrastructure control toward managerial influence through agents (all actions via agents; UI only for human decision support)
  • Operational scale examples:
    • Repo operations: +80%
    • Automated checks/builds also “growing significantly”

• End-to-end automation potential

  • Study outcome:
    • 30–40% of PRs are “routine” (migrations, refactoring, support, post-incident)
    • ~70% of those can be automated end-to-end (zero human involvement)
  • Current automation via specialized “infrastructure engine”:
    • 10% of PRs resolved in individual units using this engine
    • Company-wide described as around ~2% (more modest but rapidly growing)
    • Expected eventual level: ~25–50% of PRs

• Analyst task automation

  • Similar infrastructure aims to automate analytics tasks:
  • Result/trajectory: ~30–50% of analytics tasks solved agentically.

---

Frameworks / playbooks explicitly or implicitly used

• Adoption-first rollout

  • Teach skill → enable access → embed agents into daily routines.

• Measurement model for productivity

  • “Time boost per activation” × “number of tool uses”
  • Baseline comparisons (AI vs non-AI) used early; later limited by near-universal adoption.

• Two-stage scaling philosophy

  • Move from tools → to “Agent X” (agent-first automation), not “DWEX” (as stated in talk)
  • Then progress toward end-to-end problem solving and self-improving loops.

• End-to-end automation engine concept

  • Dedicated infrastructure with interfaces across tracker/messenger/web to trigger agents solving full problems.

• Closed-loop optimization / autoresearch

  • A loop where an agent makes changes, checks, rolls forward/backward, and repeats to improve execution.

• Reinforcement-learning-like development cycle (agent cycles)

  • “Idea agent → Implementation agent → Testing agent” in a cycle; evolve tools/prompts based on measurable outcomes.

---

Concrete initiatives & actionable recommendations (as described)

• Train for adoption

  • Invest in training + certification profiles + “ambassador” support to remove adoption friction.

• Reduce reliance on specialists/tools

  • As universal agents catch up, retire narrow specialized tools to simplify the ecosystem.

• Restructure infrastructure for agent workflows

  • Ensure infrastructure actions occur through agents (agent-mediated ops).
  • Scale capacity for expected 10× workload.

• Automate routine PRs end-to-end

  • Build an infrastructure engine to let agents resolve migrations/refactoring/support/post-incident PRs with minimal/no human involvement.

• Implement analytical agent pipeline

  • Extend agent automation beyond engineering to analytics tasks through similar infrastructure.

• Adopt closed-loop optimization

  • Use autoresearch cycles for performance/optimization problems and generalize to product/problem domains where an execution environment and measurable objective exist.

---

Leadership & workforce transformation

• Engineering role shifts

  • From “writing software” → “building factories for software production.”
  • Engineers move toward “LED” (leader) functions:
    • Decompose tasks
    • Negotiate/coordinate
    • Delegate to agents
    • Verify outcomes
    • Own product understanding

• Broader convergence

  • Similar agent-driven skills are spreading beyond engineering into other professions (product, design/other cited categories), suggesting organizational skill homogenization around agent workflows.

• Constraints / risks (not everything is smooth)

  • Smaller teams but still dependent on individuals (vacation/on-call/support coverage issues).
  • University training vs real work skills diverge; people can’t keep up easily due to rapid change.
  • Higher parallel work increases cognitive load and fatigue.
  • Potential quality decline risk if validation/checking intensity drops (even with smart models).

---

High-level business direction stated

Yandex’s “next” emphasis is progressing from agent assistance to agent-directed execution, then to system-level requirement solving and self-improvement loops, supported by infrastructure designed for agent-mediated operations at dramatically higher scale.

---

Presenters / sources

• Presenter: Andrey Popov (Yandex)
• External source referenced: GitHub (reported doubling of commits in 4 months)