How to Get Your Org to Adopt Coding Agents (Without Shipping Garbage) - Eyal Blum, Figma

Adopting AI agents into engineering workflows requires moving beyond initial productivity gains to building sustainable practices through verification, plann...

By Sean Weldon

Abstract

The integration of AI agents into software engineering workflows presents significant organizational and technical challenges that extend beyond initial productivity gains. This synthesis examines the three-phase adoption process observed at Figma, where early 10x productivity improvements give way to scaling failures before sustainable practices emerge. The analysis identifies critical friction points including reduced developer agency, communication volume increases of 2-3x, and the emergence of best engineers as organizational bottlenecks. Key findings indicate that verification infrastructure, planning-centric workflows, and attention-aware communication protocols are essential for maintaining code quality during AI adoption. Practical solutions include Test-Driven Development (TDD) style prompting, hierarchical verification systems prioritizing deterministic analysis, and explicit AI-authorship marking. These findings provide actionable frameworks for organizations navigating the simultaneous cultural and technical transformation required for sustainable AI agent integration while preserving engineering satisfaction and institutional knowledge.

1. Introduction

The deployment of AI agents in software engineering workflows represents a fundamental transformation in how code is produced, reviewed, and maintained. While initial adoption frequently yields dramatic productivity improvements, organizations rapidly encounter challenges that threaten code quality, team cohesion, and engineering satisfaction. This synthesis examines empirical findings from Figma's AI agent adoption process, focusing on the transition from experimental usage to sustainable organizational practices.

AI agents in this context refer to autonomous or semi-autonomous systems capable of generating code, analyzing repositories, and performing engineering tasks with minimal human intervention. These systems extend beyond simple code completion tools to encompass planning, implementation, and verification capabilities. The central research question addresses how organizations can scale AI agent usage beyond initial quick wins while maintaining code quality, developer satisfaction, and institutional knowledge transfer.

The challenge of sustainable AI adoption differs fundamentally from traditional technology integration. Unlike conventional tools that augment existing workflows, AI agents potentially restructure the fundamental nature of engineering work itself. This restructuring creates second-order effects on team dynamics, communication patterns, and quality assurance processes that must be addressed systematically rather than incrementally.

The analysis proceeds through three primary domains: the adoption lifecycle and organizational friction points, technical verification and workflow strategies, and communication protocols for human-AI collaboration. These findings provide a framework for understanding both the promise and pitfalls of AI-augmented software development.

2. Background and Related Work

2.1 The Three-Act Adoption Framework

AI agent adoption follows a predictable three-phase pattern that differs substantially from traditional technology adoption curves. Act 1 consists of initial experimentation where simple tasks demonstrate 10x productivity improvements, generating enthusiasm and rapid adoption. Engineers experience immediate value from automating routine tasks, creating organizational momentum for broader deployment.

Act 2 emerges when teams attempt to scale these practices to complex problems, revealing agent limitations, increased bug rates, and erosion of trust in AI-generated outputs. This phase represents a critical juncture where organizations must resist abandoning AI tools entirely and instead invest in systematic improvements. Act 3 represents the development of sustainable practices through proper guardrails, context management, and verification infrastructure. This framework emphasizes the necessity of traversing through failure modes to achieve genuine organizational capability.

2.2 Organizational Heterogeneity in Adoption

A critical observation is that adoption proceeds unevenly across teams within the same organization. Some teams achieve full transformation while others remain experimental or skeptical. This heterogeneity is not a temporary state but a persistent organizational reality requiring accommodation. The coexistence of multiple adoption stages within a single engineering organization creates unique coordination challenges not addressed by uniform rollout strategies. Furthermore, this uneven adoption reveals important information about where AI agents provide genuine value versus where they introduce friction without commensurate benefits.

3. Core Analysis

3.1 Organizational Friction Points and Human Factors

Three primary friction points emerge during AI agent scaling that fundamentally alter the engineering experience. First, reduced developer agency manifests as engineers transitioning from active coding flow states to passive prompt-wait cycles. This shift fundamentally alters the nature of engineering work, causing job satisfaction degradation and burnout. Engineers report losing the intrinsic satisfaction of direct code creation, replaced by the frustration of iterative prompting and waiting for agent outputs.

Second, the phenomenon of best engineers as bottlenecks creates a paradoxical situation where the most capable engineers become the slowest to adopt AI tools. These engineers possess institutional knowledge and mental models of where agents fail, enabling them to prevent poor outputs but simultaneously becoming overburdened with verification responsibilities. Their skepticism stems from direct observation of failure modes that less experienced engineers may not recognize, making their concerns valuable diagnostic information rather than mere resistance to change.

Third, communication inefficiency emerges as a substantial drag on organizational velocity. Design documents, Slack messages, and emails increase to 3-4x their previous length, with message volume increasing 2-3x despite containing the same information density. This expansion creates signal detection problems where distinguishing high-quality contributions from AI-generated filler becomes increasingly difficult. The cognitive load of processing this expanded communication volume partially offsets productivity gains from code generation automation.

3.2 Verification Infrastructure and Deterministic Workflows

The analysis identifies verification investment as the highest-value intervention for sustainable AI adoption. The strategy involves shifting verification responsibilities from humans to agents where possible, then encoding proven agent workflows into deterministic flows that eliminate token consumption and ensure consistency. For example, integration of Playwright and Model Context Protocol (MCP) enabled agents to explore codebases autonomously rather than requiring human navigation, representing a major productivity unlock.

The testing pyramid framework provides a hierarchical approach to verification. The foundation consists of deterministic analysis including linting, compiler checks, and unit tests that should handle the majority of verification tasks. The middle layer involves agent-based review of architectural standards and code patterns. The apex reserves human review exclusively for functionality decisions and strategic questions about whether the correct problem is being solved. This hierarchy ensures that expensive human attention focuses on areas requiring judgment and domain expertise rather than mechanical verification.

Test-Driven Development (TDD) style prompting emerges as a superior approach to agent task specification. Setting goals first and having agents strive toward them produces better results than generating code and fitting tests afterward. This approach provides agents with clear success criteria and prevents drift during implementation. Each development phase must have independent validation gates and acceptance criteria to prevent cascading failures where errors in early phases compound in later work.

3.3 Planning-Centric Workflow Architecture

A fundamental insight involves shifting engineering craft from code-writing to detailed planning, thereby restoring developer agency and job satisfaction. The planning-first workflow requires significant upfront investment - exemplified by spending one week writing comprehensive plans with decisions fully specified and reviewed before agent implementation begins. Plans should commence with a 'why' section to prevent agent drift and include executive summaries providing context for agent processing.

Plans must be decomposed into small, independently verifiable components sized to fit within a single pull request review session - the coffee-break test provides a practical heuristic for appropriate sizing. This decomposition enables parallel work by sub-agents and ensures that review cycles remain manageable. The structure includes executive summaries, phased breakdowns with detailed sub-sections, and explicit acceptance criteria for each component.

Empirical results demonstrate substantial productivity gains: six weeks of coding work compressed to one week of planning plus overnight agent implementation, yielding a 5x speedup including review cycles. This compression occurs because the planning phase forces resolution of ambiguities and edge cases that would otherwise emerge during implementation, reducing iteration cycles and rework.

3.4 Attention-Aware Communication Protocols

The recognition that human attention is a scarce resource in the AI era necessitates explicit protocols for distinguishing AI-generated from human-written content. Pull request descriptions should commence with human-written summaries of functionality, followed by AI-generated implementation details. This structure enables reviewers to quickly assess intent before diving into mechanical details.

Clearly distinguishing authorship allows readers to calibrate scrutiny levels appropriately and maintains trust within engineering organizations. An incident where AI-analyzed feedback was distributed without clear attribution damaged trust with senior engineers, demonstrating the reputational risks of ambiguous AI usage. Transparency about AI involvement in Slack messages, emails, and documents helps set appropriate expectations and prevents erosion of interpersonal trust.

The communication protocol extends beyond simple labeling to encompass strategic decisions about when AI generation adds value versus when it creates noise. Not all content benefits from AI expansion; concise human communication often proves more effective than verbose AI-generated alternatives.

4. Technical Insights

4.1 Implementation Patterns and Tooling

The integration of Playwright with Model Context Protocol (MCP) represents a critical technical enabler, allowing agents to autonomously explore codebases rather than requiring human-guided navigation. This capability eliminates a major bottleneck in agent productivity and enables more sophisticated analysis workflows.

TDD-style prompting produces measurably better outcomes than code-first approaches. By specifying success criteria before implementation, agents receive clear optimization targets and avoid the common failure mode of generating code that technically runs but fails to solve the intended problem. This approach maps naturally to formal verification concepts where specifications precede implementation.

4.2 Workflow Flexibility and Adoption Tactics

Practical adoption benefits from meeting engineers where they are rather than mandating uniform workflows. Normalizing everyday AI use reduces friction and allows organic discovery of effective patterns. Tagging agents in Slack messages for specific tasks with loop closure in threads proves highly effective, providing lightweight automation without requiring complex infrastructure.

Non-passive-aggressive agent invitations - phrased as "Let's try if the agent can get this" - help skeptics experience positive outcomes without feeling coerced. This framing positions AI as an experimental tool rather than a mandated replacement for human work. Sophisticated automation workflows can layer atop simple, accessible practices once basic patterns prove valuable.

4.3 Limitations and Ongoing Challenges

The automation story remains incomplete, particularly regarding cloud agent usage given system dependencies and security requirements. Organizations continue experimenting with the appropriate balance between local and remote agent execution. Additionally, the simultaneous cultural and engineering transformation creates coordination challenges that lack established solutions. Figma's continued learning despite shipping AI products externally demonstrates that even organizations at the frontier face substantial unknowns.

5. Discussion

The findings reveal that sustainable AI agent adoption requires addressing human factors as rigorously as technical implementation. The shift from coding to planning represents not merely a workflow change but a fundamental reconceptualization of engineering craft. This transformation restores agency by repositioning engineers as architects and validators rather than passive prompt operators. The success of planning-centric workflows suggests that AI agents excel at mechanical translation of well-specified intent rather than ambiguous problem-solving.

The emergence of best engineers as bottlenecks provides valuable diagnostic information about where verification infrastructure remains inadequate. Rather than viewing skepticism as obstruction, organizations should treat it as a roadmap for improvement. Engaging skeptics in solution design rather than forcing adoption accelerates the development of genuinely useful tools that address real workflow gaps.

The attention-aware communication protocols address a fundamental challenge of human-AI collaboration: maintaining trust and signal quality in an environment of increased content volume. The 2-3x increase in communication volume with 3-4x longer messages threatens to overwhelm human processing capacity, making explicit authorship marking essential for sustainable collaboration. This finding has implications beyond software engineering for any domain integrating AI-generated content into human workflows.

Future research should examine the long-term effects of planning-centric workflows on engineering skill development and institutional knowledge transfer. Questions remain about whether junior engineers develop appropriate mental models when separated from implementation details, and how organizations maintain architectural coherence across AI-generated codebases. Additionally, the heterogeneity of adoption patterns warrants further investigation to understand which organizational and technical factors predict successful versus problematic AI integration.

6. Conclusion

This synthesis demonstrates that sustainable AI agent adoption in software engineering requires a coordinated approach spanning verification infrastructure, workflow redesign, and communication protocols. The three-act adoption framework provides organizations with realistic expectations about the trajectory from initial enthusiasm through scaling challenges to sustainable practices. Key technical interventions include hierarchical verification systems prioritizing deterministic analysis, TDD-style prompting for clear success criteria, and planning-centric workflows that restore developer agency.

The practical takeaways emphasize meeting teams where they are, treating skepticism as diagnostic information, and investing heavily in verification infrastructure before scaling agent usage. Organizations should expect communication volume increases of 2-3x and implement attention-aware protocols that clearly distinguish AI-generated from human-written content. The 5x productivity gains observed in planning-first workflows demonstrate that proper implementation can deliver substantial value while maintaining code quality and engineering satisfaction.

Future work should focus on long-term skill development implications, institutional knowledge transfer mechanisms, and the evolution of architectural coherence in AI-augmented codebases. As AI agents become more capable, the frameworks presented here provide a foundation for navigating the ongoing transformation of software engineering practice while preserving the human elements that drive innovation and maintain organizational effectiveness.

Sources

About the Author

Sean Weldon is an AI engineer and systems architect specializing in autonomous systems, agentic workflows, and applied machine learning. He builds production AI systems that automate complex business operations.

LinkedIn | Website | GitHub