# What the Top Companies Do Differently With AI Coding - Martin Harrysson & Prakhar Dixit, McKinsey

## Executive summary

Top companies leveraging AI agents in software development are moving beyond modest gains, with the top 20-25% reporting 2x end-to-end improvements. Achieving this requires systemic changes across the entire software development lifecycle (PDLC), shifting from traditional two-week sprints to autonomous, agent-driven, Kanban-style workflows. Key focus areas include standardizing artifacts, building a robust semantic context layer, restructuring teams into smaller, merged-role pods, and proactively managing governance and quality gates throughout the process.

## Key takeaways

- Top performers see 2x end-to-end improvements using AI agents.: While most large companies report modest gains, the top 20-25% are achieving significant, real 2x end-to-end improvements by adopting new operational patterns.
- Operating models must shift from sprints to autonomous, short cycles.: The traditional two-week sprint model is obsolete in the agentic PDLC. Clients are moving toward Kanban-style, one-day/two-day workflows where agents run the night shift, and humans validate the output.
- The bottleneck shifts from Build/Test to Requirements and Release.: As build and test stages accelerate, the primary bottlenecks are shifting upstream (requirements gathering) and downstream (release planning and compliance).
- The critical technical edge is the semantic context layer.: The tooling is largely solved; the competitive advantage lies in how an organization organizes its stack around tools, specifically by building a semantic layer that captures and feeds institutional knowledge (IP, context) to the agents.

## Technical details

- Operating Model Shift: Traditional two-week sprints are being replaced by autonomous, Kanban-style, one-day/two-day workflows. This requires defining clear guardrails for human review and allowing agent loops to proceed autonomously based on confidence levels.
- Artifact Standardization: It is critical that artifacts (e.g., PRDs, specs) used to feed into agent loops are predefined and standardized across the organization to ensure consistency, especially in large enterprises.
- Team Structure and Roles: Teams are shrinking into smaller pods (2-4 people) with merged responsibilities. Roles are evolving toward standardized definitions, such as 'definer' and 'builder' roles, as manual activities are automated.
- Context and Governance: The primary technical challenge is ensuring agents have access to institutional context (SME knowledge, logs, customer emails). Furthermore, a 'trust layer' for governance, security, and stability is becoming increasingly vital.
- Quality Management: As code generation speeds up, quality can drop. Mitigation requires coaching developers and testers to focus on the skill set of reviewing AI output and curating 'golden examples' to define what 'good' code looks like at every checkpoint (PRDs, code, etc.).

## Practical implications

- Re-evaluate the SDLC process flow, moving away from fixed-length sprints toward continuous, agent-orchestrated, short-cycle workflows.
- Standardize and centralize core development artifacts (e.g., PRDs, specifications) to serve as consistent context for AI agents.
- Restructure teams into smaller, cross-functional pods (2-4 members) and redefine roles to eliminate redundant manual tasks.
- Prioritize building a semantic layer that captures and makes institutional knowledge accessible to AI agents, rather than relying on individual Subject Matter Experts (SMEs).
- Implement governance and quality gates at every checkpoint (from requirements to code merge) to manage the risks associated with increased code generation speed.

## Topics

AI Agents, Software Development Lifecycle (PDLC), Operating Model Transformation, Semantic Layer, DevOps/Build Engineering, Governance and Compliance, McKinsey, AIE CODE summit

Source: https://www.youtube.com/watch?v=21tuT-e4NVo
