How AI Agents Are Changing Software Development Forever

The 2026 AI Engineer World’s Fair revealed how AI is reshaping software engineering. The focus was on building systems where AI agents don’t just help write code—they help maintain the system itself. This concept, called autoresearch, lets agents build loops to keep improving and verifying their own work.

Thariq Shihipar from Anthropic described this shift well. He said, “The models are grown, not developed.” Instead of crafting models upfront, engineers learn alongside the model as it evolves. This means designing tools that agents can understand and use properly. The way inputs and outputs are structured changes how well an agent performs.

One example Shihipar gave was a tool called “AskUserQuestion.” His team tried three versions with the same goal but saw very different agent behaviors. This shows tools must be designed with agents in mind, not just humans. Shihipar added that the failure often lies in the tool’s interface, not the AI model itself.

Many speakers stressed the importance of keeping humans involved. Addy Osmani argued that the “outer loop” should stay with people. He explained, “That inner loop is capability. The outer loop is agency.” In other words, AI can handle tasks inside loops, but humans must decide what loops to build and verify results. He urged engineers to spot repetitive tasks and turn them into loops for agents to handle.

Paul Bakaus, the designer of Impeccable, rejected the idea of full automation. He said, “There is no auto, and there will be no auto.” His view is that agents should do about 80% of the work. Humans add the final 20% to steer outcomes and ensure quality. Bakaus summed it up: “You can’t one-shot design.” It takes trial, error, and collaboration between people and AI.

Polarization Between Human Experts and Agent Users

Geoffrey Litt from Notion warned about a growing divide. He tweeted, “Factories is a depressing vision of the future, metaphors matter.” He predicts a split between two groups: those who deeply understand their code and keep innovating, and those who delegate understanding to agents and risk being replaced. “You can learn what the agent is doing to make sure you can be an active participant in the creative process,” he said.

This divide touches on trust and expertise. Nicole Brichtova at Google pointed out that expert human judgment is still critical, especially in creative media. “Somebody who has honed a craft has a very different level of expertise,” she said. Model developers will need to work closely with people who bring creative viewpoints. She added, “It ends up being us. It ends up being the modeling teams.”

AI’s Growing Role in Code and Verification Challenges

AI-generated code is a major part of today’s software. Daksh Gupta, CEO of Greptile, shared data showing AI-generated pull requests rose from under 1% in early 2025 to 27.6% by April 2026. These AI PRs tend to be about 20% larger than human-written ones. Interestingly, their reversion rates are about the same or even lower for bigger AI PRs.

Despite this growth, trust is a concern. A survey by Sonar found 96% of developers don’t fully trust AI-generated code to be correct. Less than half always verify AI-assisted code before committing it. Yet, AI code now makes up 42% of all commits and is expected to reach 65% by 2027.

To handle this, verification frameworks must evolve. Ameya Bhatawdekar and Vinoth Govindarajan emphasized the need for new evaluation methods. Agents act in multi-step, path-dependent ways. Static, one-step tests no longer cut it. Micro-sandboxes now securely run agent-generated code to prevent security issues. Credential masking protocols limit agent access to sensitive data and reduce prompt injection risks.

Tariq Shaukat, CEO of Sonar, said, “In the Land of AI Agents, the Verifiers Are King.” This highlights how crucial it is to check AI work carefully. Addy Osmani noted, “The model that wrote the code is way too nice grading its own homework.” Human oversight remains essential.

The New Engineering Frontier: Harness Engineering and System Design

The industry is moving toward multi-agent systems that work across entire code repositories, not just single files. Uber’s uReview system was highlighted as an example. This shift requires formal harness engineering—building wrappers around models to enforce rules, manage state, and stop infinite loops.

Context engineering is also key. It treats the agent’s context window like a dynamic memory cache. Techniques like semantic compression and structured retrieval help agents remember what matters. Agents also navigate GUIs through visual inputs rather than APIs, which adds complexity.

Conference keynote speaker Addy Osmani described how systems should discover work, run in isolated contexts, verify outputs with different agents, save state, and connect with external tools. This layered design helps keep AI behaviors reliable and predictable.

Barr Yaron from Amplify Partners shared that over 70% of organizations now use three or more AI models. Agent framework adoption nearly doubled from 9% to 18% in one year. Most teams (89%) have some form of agent observability, and 62% track detailed step-level agent actions.

Garry Tan of Y Combinator closed the fair, likely sharing insights about building AI-native companies and teams. The broader themes included agentic commerce, AI in regulated industries, local inference, and knowledge graph integration.

Overall, the takeaway is clear. AI agents are reshaping software development. But this future is not about full automation. It’s about smart partnerships where humans guide, verify, and build on AI capabilities. Engineers should focus on automating repetitive tasks, designing solid systems, and staying deeply involved in the process.