Scaling AI Agents: From Editor to Infrastructure
James Everingham of guild.ai shares how Meta’s DevInfra team shifted from AI autocomplete to agentic infrastructure. Learn why centralized control planes are essential for governing agent workflows, reducing onboarding time, and eliminating code freezes through organic, challenge-driven adoption.
The Shift from Autocomplete to Agentic Infrastructure
Engineering leaders are increasingly finding that AI tools focused solely on code completion yield diminishing returns. James Everingham, former head of DevInfra at Meta and CEO of guild.ai, argues that the next phase of AI adoption requires treating agents as sentient fabric within the Software Development Life Cycle (SDLC). At Meta, this shift transformed internal developer experience, with agents eventually submitting 50% of all code diffs. The core strategy involves moving AI capabilities closer to source control, the canonical source of truth, to intercept tool connections and invoke agents for broader infrastructure tasks.
Measuring True Productivity
A critical barrier to AI adoption is the lack of clear metrics. Traditional measures like lines of code are ineffective, as high-productivity engineers often delete more code than they write. Instead, organizations should focus on operational metrics such as Time to First Diff (TTFD). At Meta, an onboarding agent that made the codebase 'sentient' significantly reduced TTFD, reclaiming substantial engineering hours. This demonstrates that AI’s highest value lies in reducing friction in complex, large-scale codebases rather than merely accelerating typing speed.
Organic Adoption Through Challenges
Top-down mandates for AI usage often fail because developers must earn trust in new tools. Everingham recommends issuing high-level business challenges, such as eliminating code freezes or creating self-healing systems, rather than dictating specific tools. These challenges force teams to explore AI solutions organically. When engineers see peers succeeding with these tools, adoption spreads naturally. This approach fosters a culture of experimentation where engineers can fork and improve existing agents in a centralized platform, similar to an internal app store.
The Need for Agentic Control Planes
As agents scale from individual 'single-player' modes to enterprise-wide 'multiplayer' systems, governance becomes critical. Organizations need a control plane to manage agent access, logging, costs, and security. This infrastructure ensures that agents operate within defined guardrails, satisfying security and compliance requirements. Without centralized governance, the risk of fragmented, unmanaged agent deployments increases, leading to potential security vulnerabilities and cost overruns. The future of engineering infrastructure involves thousands of agents, requiring robust orchestration and monitoring capabilities to maintain stability and efficiency.
Key insights
-
AI productivity gains are limited when confined to code completion. Significant leverage is found by integrating agents into source control and infrastructure layers.
Impact: Enables systemic automation of SDLC tasks, moving beyond individual developer speed to organizational throughput.
-
Centralized platforms for agent discovery and forking accelerate innovation by allowing engineers to build on existing solutions rather than creating from scratch.
Impact: Reduces time-to-value for new AI workflows and fosters a collaborative engineering culture.
-
Traditional metrics like lines of code fail to capture AI impact. Operational metrics like Time to First Diff (TTFD) better reflect true productivity improvements.
Impact: Provides leadership with actionable data to justify AI investments and track onboarding efficiency.
-
Organic adoption is driven by high-level business challenges rather than tool mandates. Engineers adopt AI when it is necessary to solve complex problems like code freezes.
Impact: Increases buy-in and reduces resistance to new technologies by aligning AI usage with business goals.
-
Enterprise-scale agent deployment requires a control plane for governance, security, and cost management to prevent fragmentation and risk.
Impact: Ensures compliance and security while enabling the safe scaling of agentic workflows across the organization.
Action items
-
Audit current AI tools to identify opportunities for moving capabilities from the editor to source control and CI/CD pipelines.
Impact: Unlocks deeper integration points for automation, increasing the scope of AI impact beyond code writing.
-
Establish a centralized repository for internal AI agents and workflows, allowing engineers to discover, fork, and improve them.
Impact: Accelerates innovation by leveraging collective intelligence and reducing redundant development efforts.
-
Define and track operational metrics such as Time to First Diff (TTFD) and code freeze duration to measure AI impact.
Impact: Provides clear, business-relevant data to evaluate ROI and identify areas for further optimization.
-
Issue high-level business challenges, such as eliminating code freezes, to drive organic AI adoption among engineering teams.
Impact: Encourages creative problem-solving and natural tool adoption without the friction of top-down mandates.
-
Implement an agentic control plane to manage agent access, logging, and costs across the organization.
Impact: Ensures security, compliance, and cost visibility as agent usage scales, mitigating operational risks.
Quotes
“if you actually move some of this new agentic behavior closer to your source control system as part of the infrastructure, you get a lot more impact”
“the measure of a successful initiative or a successful platform is when the ideas are coming from your customers and they're able to build on it”
“if 2025 was the year of ibe coding, like 2026 is sort of the year of agents”