Coding agents need a well-designed harness, clean context, reliable verification, and infrastructure that lets them work across long tasks.
2
AI adoption produces uneven results, so companies need to change team structures, workflows, training, incentives, and measurement instead of only buying tools.
3
The companies getting the most from AI treat software development as a system that includes code quality, documentation, review, security, and business outcomes.
Summary
This leadership track is a collection of practical arguments about how AI changes software organizations. Anthropic describes agent platforms around model capabilities, context management, memory, context editing, code execution, and skills. Replit argues that autonomy should be scoped to the user's goal, with verification and context management supporting nontechnical users. OpenAI explains why the harness around a model is difficult to build and why it may become a reusable abstraction layer. McKinsey and several enterprise speakers focus on operating-model changes, smaller teams, continuous planning, measurement, and gradual rollout. Other talks warn that faster code generation can increase rework, defects, review load, and security risk. The later sessions discuss proactive agents, AI-native products, compensation by output, and compounding engineering, where each completed feature improves the next one. Across the day, the speakers repeatedly connect AI results to the surrounding engineering system rather than to model capability alone.
Agent performance depends on capabilities, context, and a computer they can operate
Katelyn Lesse describes Anthropic's platform through three functions: expose Claude's capabilities, manage its context window, and give Claude a computer to use. The examples include configurable thinking budgets, built-in and custom tools, MCP connections, memory, and context editing that clears old tool results. Anthropic combined memory with context editing and saw a 39% performance increase on its internal evaluations. Claude Code can run code in a secure sandbox, while agent skills add scripts, instructions, and resources for tasks such as applying a web design system. Lesse argues that agents become more useful when they can work autonomously inside a controlled environment instead of relying only on a prompt and a fixed tool list.
Autonomy should be scoped to user goals and supported by verification
Michele Catasta separates supervised autonomy, where a technically skilled user remains ready to intervene, from Replit's goal of making autonomous coding usable by nontechnical people. Long runtime alone is not the measure. A narrowly scoped task can be autonomous and fast, while a broad task may run for hours. Replit's approach rests on model capability, verification, and context management. Catasta says agents build many 'painted doors' without testing, such as buttons whose handlers are missing or pages backed by mock data. Replit uses autonomous testing with database access, logs, API calls, browser interaction, and Playwright code. Catasta says this approach is roughly an order of magnitude cheaper and faster than computer use, while also producing reusable regression tests.
A coding agent is a user interface, a model, and a harness
Bill Chen and Brian Fioca reduce a coding agent to three parts: the interface, the model, and the harness. The harness is the interface layer that supplies prompts and tools, maintains the core loop, handles multiple turns, and lets the model work on code. They describe the difficult engineering around parallel tool calls, thread merging, sandboxing, permissions, port management, compaction, MCP, and images. They also argue that model training creates habits, such as planning, gathering context, testing, and checking work. A harness that over-instructs a model can make it slower or less effective. Codeex is presented as a model-and-harness combination available through a VS Code plugin, CLI, cloud environments, chat, Slack, GitHub, and an SDK.
AI adoption requires new workflows, roles, and measurement
Martin Harrysson and Natasha Maniar say enterprise teams are seeing individual productivity gains while company-wide improvements often remain limited. Their explanation is that review, collaboration, work allocation, and technical debt have not changed with the faster code generation. In their survey of about 300 enterprises, stronger performers were seven times more likely to have AI-native workflows and six times more likely to have AI-native roles. They describe continuous planning, spec-driven development, one-pizza pods of three to five people, and product builders who orchestrate agents with full-stack fluency. They also argue for measuring inputs, delivery, quality, developer experience, resilience, and economic outcomes together. In one bank intervention, agent consumption increased by over 60 times and code mergers increased by 51%.
AI productivity depends on codebase hygiene and careful outcome measurement
Yegor Denisov-Blanch presents a study comparing 46 teams using AI with 46 similar teams that were not. The median productivity gain for that cohort was about 10% in July, while the gap between stronger and weaker teams was widening. Token use had only a loose correlation with productivity. An experimental environment-cleanliness index, based on tests, types, documentation, modularity, and code quality, showed a stronger relationship with AI gains. He warns that unchecked AI use can increase codebase entropy and cause engineers to reject or heavily rewrite generated output. In a case study, pull requests rose 14%, but code quality fell 9%, rework increased 2.5 times, and effective output did not meaningfully increase. His conclusion is that PR counts alone can make a negative result look positive.
Proactive agents should observe work and intervene at the right time
Kath Korevec frames proactive agents around the mental load of monitoring asynchronous work. Humans execute tasks serially, and context switching can consume a large share of productive time. Google Labs' Jules is intended to observe code changes and workflows, learn user preferences, and suggest or perform work before the user asks. Korevec describes three levels: fixing issues such as missing tests and unsafe patterns, learning the project's broader context, and understanding consequences across code, design, analytics, telemetry, and conversion data. Jules includes memory, a critic agent, Playwright-based verification, a to-do bot, best-practice suggestions, environment setup, and just-in-time context. A planned proactivity view indexes a repository, ranks possible tasks by confidence, and lets the user inspect the rationale before starting work.
Quality and security must be designed into the AI development loop
Itamar Friedman argues that code generation has a productivity ceiling unless companies add agentic quality workflows across the software development life cycle. He describes more generated code, more pull requests, and more review time, with 67% of surveyed developers expressing serious quality concerns. His proposed response includes automated quality gates, intelligent code review, testing, living documentation, and a context engine containing code, version history, pull-request history, organizational logs, standards, and best practices. The Browser Company gives a related security example: an AI browser can access private data, read untrusted content, and communicate externally. Prompt injection defenses therefore need both technical measures and user confirmations before actions such as filling forms, sending emails, or scheduling events.
AI-native organizations compound knowledge through explicit engineering practices
Dan Shipper describes Every, where four software products are built by 15 people and 99% of the code is written by agents. He calls the process 'compounding engineering': plan, delegate, assess, and codify. Codify means turning lessons from planning, implementation, testing, and review into prompts, commands, subagents, and shared instructions. This lets developers reuse tacit knowledge across repositories, makes new hires productive sooner, and allows people to work across different stacks without standardizing every language. Shipper also says AI enables a demo culture, parallel work, and fractured-attention contributions from managers. His broader point is that reaching full adoption changes the way the organization collaborates, rather than only making individual coding tasks faster.
"Keeping the right context in the window at any given time is really, really critical to getting the best outcomes from Claude."08:19
Who should watch
You lead an engineering organization that has adopted coding agents but cannot tell whether delivery speed, quality, or rework is improving.
Your team is building an agent product and needs concrete guidance on harness design, context management, verification, sandboxes, memory, and proactive behavior.
You are responsible for introducing AI inside a large or risk-sensitive company and need examples of staged rollouts, new team structures, training, governance, and outcome measurement.
