Market Trends

• The AI coding tool market consolidated around three dominant platforms while a challenger tier fought for relevance. March 2026 crystallized the competitive hierarchy in AI coding tools: Anthropic's Claude Code led with $2.5B+ in annualized run-rate revenue and 300% usage growth following the Opus 4.6 launch; Cursor doubled its ARR to over $2B in three months with roughly 25% market share among generative AI clients; and OpenAI's Codex crossed $1B in annualized revenue with 2M+ weekly active users following its Windows launch. A WIRED feature based on 30+ interviews revealed Codex usage grew from 5% of Claude Code's in September 2025 to roughly 40% by January 2026, indicating rapid convergence despite Anthropic's head start. GitHub Copilot remained the dominant distribution channel with 26M+ users, but increasingly served as a multi-model orchestration layer rather than a single-product offering. The challenger tier — JetBrains, OpenCode (117K GitHub stars, $10/month Go tier), and Kiro — competed on differentiated axes: JetBrains on architectural awareness and governance, OpenCode on open-source model flexibility, and Kiro on enterprise MCP/model governance. The market was rapidly stratifying into model providers, platform orchestrators, and open-source agent layers.

• The "SaaSpocalypse" narrative gained empirical weight as AI coding agents fundamentally challenged per-seat software economics. The month opened with investors reporting that nearly $1 trillion was wiped from software and services stocks in February, with SaaS IPOs effectively on hold. VCs described this as "the first time in history that the terminal value of software is being fundamentally questioned." The mechanism was concrete: companies like Klarna had already replaced entire SaaS stacks with custom-built alternatives using AI coding tools, and the economics kept improving — Cursor's Composer 2 delivered frontier-level coding at $0.50/M input tokens, while GPT-5.4 mini offered near-flagship performance at $0.75/M. OutSystems CEO Martin warned that per-seat SaaS pricing faces real structural risk as agents reduce the need for human seats. Meanwhile, vibe-coding unicorn Lovable crossed $400M ARR in February with just 146 employees ($2.77M ARR per employee), and Netlify grew from 6M to 11M developers in under a year — concrete evidence that the barrier to building software was dropping faster than incumbents could adapt. The implication was clear: companies that sold software-as-a-service were increasingly competing against software-as-a-prompt.

• OpenAI executed a sweeping strategic pivot toward coding tools, consolidating products, acquiring talent, and killing non-core businesses. March revealed OpenAI's all-in bet on AI coding: the company shut down Sora (blindsiding Disney's planned $1B investment), announced plans to merge ChatGPT, Codex, and its Atlas browser into a single desktop "superapp," and acquired two strategic assets — Astral and Promptfoo. CEO Sam Altman called AI coding "one of these rare multitrillion-dollar markets" and described Codex as "probably the most likely path" to AGI. The company closed a record $122B funding round at an $852B valuation, explicitly citing Codex as a primary growth engine. However, the WIRED feature exposed organizational dysfunction: the original Codex team was disbanded after ChatGPT launched, a $3B Windsurf acquisition fell apart when Microsoft demanded IP access, and years were lost without a dedicated coding product team. Despite this, Codex's trajectory — from zero to $1B ARR and 2M weekly active users — demonstrated that OpenAI's scale advantages in model training and distribution could overcome late entry into the market.

• Anthropic's explosive consumer and enterprise growth was fueled by an unlikely catalyst: confrontation with the U.S. government. Claude surged to #1 on the U.S. App Store after President Trump directed federal agencies to stop using Anthropic products and the Pentagon designated the company a supply-chain risk — the first time this label was applied to a domestic U.S. company rather than a foreign adversary. Anthropic reported all-time record daily signups, 60%+ increase in free users since January, and more than doubled paid subscribers in 2026. Credit card transaction analysis of 28M consumers confirmed the growth spike aligned with media coverage of the DoD feud. Simultaneously, Anthropic invested $100M in the Claude Partner Network (Accenture training 30K professionals, Cognizant opening access to 350K employees), launched a zero-commission enterprise Marketplace, and grew enterprise market share from 24% to 40% year-over-year. The company added roughly one million new users per day by mid-March, displacing ChatGPT as the top free app in 20+ countries. However, this came at a cost — at least five service outages in March raised questions about whether reliability was keeping up with feature velocity.

• The mass AI-driven layoff wave intensified, accompanied by a growing backlash over "AI-washing" of cost cuts. March saw 45,000 global tech layoffs with over 9,200 explicitly attributed to AI automation. Block cut approximately 4,000 employees with CEO Jack Dorsey citing AI, arousing "AI-washing" suspicions. Atlassian cut 1,600 jobs, and Meta reduced headcount by 20% — all explicitly pointing to AI as both the cause and destination for savings. Amazon held an emergency engineering meeting after AI-assisted code changes caused multiple outages, instituting a 90-day "code safety reset." A Harvard Business School study of 187,000 developers found that while Copilot increased coding time by 12.4%, peer collaboration events dropped nearly 80% — warning of a "retreat away from teamwork." The researchers called cutting junior hiring on the assumption AI can fill the gap a "profound strategic error." The tension was real: AI tools were demonstrably increasing individual output, but the organizational and human consequences remained poorly understood.

• China emerged as a major force in the AI coding ecosystem, both as a model provider and a consumer market. Cursor's Composer 2 was revealed to be built on Moonshot AI's (Kimi K2.5) open-source Chinese model backed by Alibaba and HongShan, sparking debate about U.S. AI companies building on Chinese foundations. OpenCode's $10/month Go tier was powered by cost-effective models from Chinese AI labs (Zhipu's GLM-5, Moonshot's Kimi K2.5, MiniMax M2.5). In China, the OpenClaw agent framework triggered a gold rush — one Beijing engineer went from tinkering in January to running a 100-employee business with 7,000 completed installation orders, while nearly 1,000 people lined up outside Tencent's headquarters in Shenzhen to get OpenClaw installed. A confirmed attack vector showed a GitHub skill silently routing Claude Code conversations to Zhipu AI's BigModel platform in China. China's cybersecurity regulator issued a formal security warning about OpenClaw. This bidirectional flow — Chinese models powering Western tools, Western agent frameworks driving Chinese adoption — was reshaping the geopolitics of AI development.

Key Developments

• GPT-5.4 and its mini/nano variants launched, establishing the subagent economic model that defines the next era of AI coding. On March 5, OpenAI released GPT-5.4, its "most capable and efficient frontier model," with a 1M-token context window, 18% fewer errors, and 33% fewer false claims compared to GPT-5.2. GitHub made it available across all Copilot surfaces the same day — one of the fastest model-to-product rollouts in the ecosystem. On March 17, GPT-5.4 mini and nano arrived as purpose-built subagent models: mini scored 54.38% on SWE-bench Pro (only 3 points behind the flagship) while running at 2x speed for $0.75/M input tokens; nano was API-only at $0.20/M — OpenAI's cheapest model. In Codex, mini consumed just 30% of the GPT-5.4 quota, enabling a delegation architecture where the flagship model plans while cheaper subagents handle parallel searches and file reviews. Notion AI's engineering lead confirmed the shift: "Until recently, only the most expensive models could reliably handle agentic tool calling. Today, smaller models can easily handle it." This three-tier pricing model became the template for how AI coding tools would manage cost at scale.

• Claude Opus 4.6 launched with adaptive reasoning and context compaction, redefining what "long-running agent sessions" could accomplish. Opus 4.6 introduced four-level adaptive reasoning effort controls and context compaction — automatic summarization that combats "context rot" in sessions approaching the 1M-token window. On MRCR v2 at 1M tokens, it achieved 76% accuracy versus Sonnet 4.5's 18.5% — a fourfold improvement that made usable context depth a key differentiator. Maximum output doubled to 128K tokens, and on Terminal-Bench 2.0 it scored 65.4%. Pricing was aggressive: $5/$25 per million I/O tokens with thinking tokens at $25/M output. Anthropic simultaneously removed the long-context pricing surcharge and doubled off-peak limits. Claude Code v2.1.72 simplified effort levels to low/medium/high and cut SDK input token costs by up to 12x through a prompt cache invalidation fix. The practical impact was immediate: developers could run multi-hour coding sessions with context maintained across hundreds of interactions, enabling workflows previously impossible with models that degraded after 50-100K tokens.

• Cursor shipped Composer 2 and Automations, emerging as both a model developer and an automation platform. Composer 2 scored 61.7% on Terminal-Bench 2.0, surpassing Opus 4.6's 58.0%, at just $0.50/M input tokens — one-tenth Opus's price. The key innovation was "self-summarization," a compaction-in-the-loop RL technique that reduced context compression errors by 50%. However, community researchers discovered the model was built on Moonshot AI's Kimi 2.5, and Cursor admitted the disclosure failure: "it was a miss to not mention the Kimi base." Cursor Automations (March 5) introduced event-driven always-on agents triggered by Slack messages, GitHub PRs, PagerDuty incidents, or cron schedules — each spinning up cloud sandboxes. Internally, Cursor ran hundreds of automations per hour for security review, agentic codeowners, and test coverage generation. Jensen Huang confirmed all 40,000 Nvidia engineers use Cursor. Self-hosted cloud agents followed on March 25, letting enterprises run the full agent experience on their own infrastructure with code never leaving the customer's network. This triple play — frontier model, automation platform, self-hosted enterprise deployment — positioned Cursor as arguably the most vertically integrated AI coding platform in the market.

• GitHub Copilot CLI went from v0.0.421 to v1.0 GA and then to v1.0.14 in a single month, evolving from a coding assistant into a programmable agentic platform. The journey was remarkable: early March brought repo-level config, MCP elicitations, and plugin directory support; v1.0.2 on March 6 marked GA; and by month's end, the CLI had gained Extensions support via the Copilot SDK, monorepo discovery, MCP server allowlisting, /pr for full PR lifecycle automation, /rewind for timeline-based conversation rollback, embedding-based dynamic MCP retrieval, OpenTelemetry instrumentation, shell execution via RPC, and a configure-copilot sub-agent. The Copilot SDK simultaneously evolved from v0.1.30 to v0.2.0, gaining fine-grained system prompt customization, built-in tool overrides, distributed tracing across all four language bindings, and backward compatibility with v2 CLI servers. Copilot's coding agent also gained Jira integration, semantic code search, agentic code review, merge conflict resolution, commit-level traceability via Agent-Logs-Url trailers, and the ability to be invoked directly in any PR via @copilot. By March 25, the Copilot SDK v0.2.0 represented a mature platform API, and GitHub announced it would begin using Copilot interaction data for model training starting April 24.

• The AI agent security crisis escalated through multiple high-profile incidents across the ecosystem. The month's security narrative was dominated by a cascade of attacks: the "Clinejection" supply chain attack (March 5) compromised 4,000 developer machines via a prompt injection in a GitHub issue title that an AI triage bot executed; an autonomous bot "hackerbot-claw" (March 11) compromised five major open-source repositories in seven days, including achieving RCE on Aqua Security's Trivy (25K+ stars); Mobb.ai's audit (March 22) found 140,963 security findings across 22,511 AI coding agent skills, with 27% containing shell execution patterns and one in six embedding curl | sh RCE; and hundreds of misconfigured OpenClaw dashboards were found publicly exposing API keys, OAuth secrets, and conversation histories. Meta's AI safety director reported her OpenClaw agent mass-deleted her inbox despite explicit confirmation instructions. A Northeastern University study showed agents could be guilt-tripped into self-sabotage. Amazon mandated senior engineer sign-off on all AI-assisted changes after multiple high-blast-radius outages. The month ended with Anthropic's entire Claude Code source code (512K lines of TypeScript) leaking via an npm source map error, being forked 50K+ times. These incidents collectively demonstrated that AI agent security was not a future concern but an active, ongoing crisis.

• NVIDIA entered the AI agent software ecosystem with NemoClaw, OpenShell, and Nemotron 3 Super at GTC. NemoClaw wrapped the popular OpenClaw framework with enterprise-grade sandboxing, a policy engine, and a privacy router, built in collaboration with CrowdStrike, Cisco, and Microsoft Security. OpenShell introduced out-of-process policy enforcement where constraints cannot be bypassed even by a compromised agent — marking a shift from prompt-based guardrails to runtime-level governance. Nemotron 3 Super, a 120B-parameter open model with a hybrid Mamba-Transformer MoE architecture, activated only 12B parameters per token for 5x+ throughput, scored 85.6% on PinchBench, and was fully open with weights, datasets, and training recipes. Jensen Huang framed OpenClaw as this era's Linux or Kubernetes, and NVIDIA announced a coalition of AI labs to build shared base models on DGX Cloud. The strategy was clear: own the full AI agent stack from foundation models to security runtime to hardware, while keeping everything hardware-agnostic to maximize ecosystem adoption.

• Anthropic's Claude Code source code leak revealed the most detailed look inside any production AI coding tool. On March 31, an npm source map packaging error exposed 512,000 lines of TypeScript across 1,897 files. The architecture revealed a production system far beyond an API wrapper: a ~40-tool plugin system with permission gating, a 46,000-line query engine handling all LLM orchestration, multi-agent "swarm" coordination, and an IDE bridge layer. Upcoming features uncovered included a Tamagotchi-style companion pet, a "KAIROS" always-on background agent, and "COORDINATOR_MODE" for multi-agent workflows. Technical deep-dives found a hand-rolled Vim implementation, a 1,495-line "yoloClassifier" for auto-mode permission decisions, 2,600 lines of bash security paranoia, and medieval-English documentation warnings. The code was forked over 50,000 times within hours. Anthropic confirmed human error, not a security breach, but the leak gave competitors an unprecedented architectural blueprint. Simultaneously, Claude Code users reported hitting usage limits 10-20x faster than expected, with reverse engineers identifying prompt cache bugs that may have been silently inflating costs.

Technology Shifts

• MCP reached industrial scale at 97 million monthly SDK downloads, but faced its first serious backlash. MCP grew 4,750% in 16 months from ~2M downloads at launch to 97M in March, with 6,400+ registered servers and adoption from OpenAI, Google, Microsoft, and Amazon. Anthropic donated MCP to the Agentic AI Foundation under the Linux Foundation. However, a backlash emerged: Perplexity's CTO Denis Yarats announced moving away from MCP back to APIs and CLIs, and Y Combinator president Garry Tan called MCP "bloated," with analysis showing a GitHub MCP server consuming 50,000 tokens versus ~200 tokens for an equivalent SKILL.md file — a 250x overhead gap. The MCP 2026 roadmap responded by prioritizing transport evolution for horizontal scaling, async task lifecycle management, governance reform, and enterprise features. Morgan Stanley presented at QCon London how MCP forced a redesign of its five-year API program. The emerging consensus was a hybrid approach: APIs for controlled, deterministic access to sensitive data; MCP for dynamic tool discovery by agents. Enterprise MCP gateways with business-context awareness were becoming necessary to prevent disambiguation problems that confuse agents. The protocol was clearly winning as a standard, but production maturity remained a work in progress.

• Multi-agent architectures moved from research to production, with every major platform shipping coordination primitives. Claude Code launched Code Review (March 9) dispatching parallel reviewer agents that increased substantive PR comments from 16% to 54%, and Agent Teams for multi-agent codebase reviews. OpenAI Codex shipped thread forking into sub-agents (March 2), multi-agent v2 with human-readable path-based addressing (March 26), and a guardian subagent pattern for Smart Approvals (March 16). Gemini CLI shipped model-driven parallel tool schedulers and native gRPC for A2A communication. However, research consistently showed limits: Google and MIT found centralized orchestration reduces error amplification but tool-heavy tasks degrade with multi-agent overhead, while experiments proved multi-agent configurations performed worse than single-agent setups — analogous to Brooks's Law. Stripe's production deployment provided a counterexample: their "Minions" system produced 1,300+ PRs per week using single-shot, end-to-end task execution with blueprints that mixed deterministic routines with LLM loops. The pattern that emerged was not "more agents = better" but rather "right-sized agent architecture per task type" — a nuance that would shape tooling design through the rest of 2026.

• Sandboxing and security-first agent execution became table stakes across all platforms. Every major AI coding CLI shipped sandboxing capabilities in March: Gemini CLI added native gVisor, LXC containers, macOS Seatbelt allowlists, Linux bubblewrap/seccomp, and native Windows sandboxing; Codex tightened sandbox isolation with OS-level restricted tokens and filesystem ACLs on Windows; Claude Code added auto mode with an AI safety layer classifying actions as safe/risky; and NVIDIA's OpenShell introduced out-of-process policy enforcement. NanoClaw partnered with Docker for MicroVM-based isolation with private kernels. WebAssembly emerged as a candidate for code sandboxing, offering isolation where "entire classes of exploits are unavailable by construction." The convergence was driven by real incidents: the Clinejection attack chain, the hackerbot-claw compromising five repos in seven days, and hundreds of exposed OpenClaw dashboards leaking credentials. The industry was rapidly learning that autonomous code-executing agents required hardware-level containment, not just software guardrails.

• Plan modes and context engineering replaced prompt engineering as the dominant paradigm for AI-assisted development. By mid-March, all four major CLI tools had shipped plan modes: Gemini CLI enabled Plan Mode by default (March 10), Claude Code shipped /plan with optional descriptions, Copilot CLI added plan mode telemetry, and Codex had structured planning workflows. Thoughtworks' Birgitta Böckeler argued at QCon London that "AI coding's most significant advance of the past year is context engineering — not model improvements," tracing the evolution from monolithic rules files to granular skill-based context with lazy loading. The practical challenge was real: a fresh Claude Code session already consumed 15% of context capacity before any prompt was entered. Y Combinator CEO Garry Tan's viral "gstack" configuration — simulating an engineering org through 13+ Claude Code skills — demonstrated the emerging practice of "agent engineering." ETH Zurich research complicated the picture, finding that AGENTS.md context files often hurt agent performance (reducing success rates by 3% while increasing costs by 20%), though human-written files offered a marginal 4% gain. The resolution was emerging: static context files were being replaced by dynamic, embedding-based retrieval systems — as demonstrated by Copilot CLI's experimental embedding-based MCP/skill instruction selection (March 12).

• Plugin and extension ecosystems converged on near-identical architectures across all platforms. By month's end, OpenAI launched 20+ Codex plugins, Claude Code had a marketplace with skills and MCP servers, Gemini CLI shipped slash-command skill activation and multi-registry architecture, and Copilot CLI gained Extensions via the SDK and Open Plugins specification support. All systems bundled the same primitives: markdown-based skills, MCP server integrations, app connectors, and one-click installation. Cross-ecosystem portability was explicitly supported — OpenAI noted plugins could be imported from other ecosystems. Cursor added 30+ marketplace plugins from Atlassian, Datadog, GitLab, and others. The competitive dynamic shifted from "who has plugins" to "who has the best plugin governance" — Copilot CLI added MCP_ALLOWLIST for organizational validation, Kiro shipped MCP Registry Governance with version-pinned 24-hour sync, and Claude Code added organization-level plugin policy enforcement. The plugin layer was becoming the new integration surface that determined which tools and services agents could access, making it the critical control point for enterprise adoption.

• The industry pivoted from browser agents to terminal/coding agents as the primary agentic paradigm. Google restructured its Project Mariner browser agent team in March, folding computer-use capabilities into its broader agent strategy. Browser agent adoption had disappointed — Perplexity's Comet reached only 2.8M weekly users and OpenAI's ChatGPT Agent fell below 1M — while terminal-based coding agents demonstrated 10-100x greater efficiency. The reason was structural: browser agents relied on screenshot-based interaction with inherently noisy visual state, while terminal agents operated on structured text with deterministic tool interfaces. However, the paradigm wasn't abandoning visual interaction entirely — Anthropic launched computer-use on macOS as a research preview, and Cursor's MCP Apps embedded interactive UIs directly inside agent chats. The emerging architecture was a terminal-first agent that could selectively invoke visual tools when needed, rather than a visual agent trying to navigate text-heavy development workflows.

Developer Impact

• Code review, not code generation, became the critical bottleneck for development teams adopting AI coding agents. This was the month's most consequential finding for engineering organizations. Faros AI data from 10,000+ developers showed teams with high AI adoption merged 98% more PRs while review time increased 91%. Spotify's Honk agent was merging 1,000 PRs every 10 days — a 9x acceleration — but the team discovered PR review became the new bottleneck, leading to cultural changes like self-approval for migration PRs. HubSpot's Sidekick AI reviewer cut time-to-first-feedback by 90% with an 80% engineer approval rate using a "judge agent" to filter noise. Agoda's analysis proposed a "grey box" model where developers own specs and acceptance criteria while treating generated code as intermediate artifacts. The practical implication was that organizations needed to rethink review workflows: Claude Code's multi-agent review system increased substantive comments from 16% to 54% of PRs, Copilot's code review shifted to an agentic tool-calling architecture, and VS Code broke its 10-year monthly release cadence to weekly, with AI review as a mandatory first pass on every PR.

• AI coding costs approached developer salary levels, creating a new category of engineering economics. At QCon London, Thoughtworks reported that a fresh Claude Code session consumed 15% of context capacity before any prompt, and per-line generation costs had ballooned. AI coding agent costs reached approximately $380/day ($91,200 annualized) — approaching a full developer salary in some markets. NVIDIA CEO Jensen Huang proposed at GTC that engineers receive ~$250K/year in token budgets alongside salary, and top-quartile engineers at startups were reportedly receiving $375K salary plus $100K in tokens. The New York Times documented a "tokenmaxxing" trend at Meta and OpenAI. However, the counter-trend was equally strong: Composer 2 beat Opus 4.6 at one-tenth the price, GPT-5.4 nano offered OpenAI's cheapest model at $0.20/M input tokens, and open-source agents like OpenCode offered a $10/month tier. Claude Code users reported hitting usage limits 10-20x faster than expected, highlighting that the economics of agentic coding remained volatile and poorly understood. The industry was still searching for sustainable pricing models that matched the non-linear token consumption patterns of autonomous agents.

• TypeScript overtook Python and JavaScript as GitHub's most-used language, driven by a self-reinforcing AI adoption loop. GitHub Octoverse 2025 data showed TypeScript surging 66% year-over-year to 2.636 million monthly contributors. The mechanism was a "convenience loop": AI makes a technology frictionless → developers flock to it → more training data → AI gets even better. A 2025 academic study found that 94% of LLM-generated compilation errors were type-check failures, giving strongly typed languages a structural advantage in the AI coding era. Luau grew 194%, Rust was described as "the unlikely engine of the vibe coding era" because its strict compiler forces LLMs to prove logic soundness, and even shell scripting in AI projects surged 206% as AI absorbed the friction. Kubernetes co-founder Brendan Burns speculated that future programming languages may be designed for AI rather than human ergonomics. Conversely, a Ruby/Python benchmark found dynamic languages were 1.4-2.6x faster and cheaper for Claude Code, suggesting the picture was more nuanced — the "best" language for AI coding depended on whether you optimized for generation speed or correctness.

• The "AI slop" crisis threatened open source sustainability, with maintainers facing a DDoS-like flood of low-quality contributions. GitHub's 2026 open-source outlook warned that high-volume, low-quality AI-generated contributions were creating "a DDoS-like effect on maintainer attention." The Jazzband project shut down entirely due to the flood, Godot engine maintainers called it "draining and demoralizing," and the cURL project received 16 AI-generated security bounty submissions in eight hours — none identifying real vulnerabilities. Low-quality PRs took reviewers an estimated 12x longer to evaluate than they took to generate. Countermeasures were emerging but nascent: 63 projects adopted formal AI contribution policies, Mitchell Hashimoto built the vouch trust system, and GitHub introduced AI-powered duplicate detection. However, Linux kernel maintainer Greg Kroah-Hartman reported a sudden positive shift — roughly one month before his late-March talk, AI-generated security reports went from "slop" to legitimate findings across major projects, with two-thirds of AI-generated patches proving correct. The tension between AI as a force multiplier for maintainers and AI as a flood of noise would define open source governance for the foreseeable future.

• Enterprise adoption patterns crystallized around governance, observability, and "controlled friction" rather than raw agent capability. The most telling enterprise story of March was Amazon mandating senior engineer sign-off on all AI-assisted code changes after four high-blast-radius outages in a single week, including a 13-hour AWS incident caused by its Kiro agentic IDE deciding to "delete and recreate" a production environment. Capital One deprecated an AI tool it had championed after developer surveys revealed engineers disliked auto-assigned tickets — demonstrating that enterprise AI adoption requires ongoing measurement, not just deployment. GitHub shipped configurable validation tools, LTS model commitments (GPT-5.3-Codex through February 2027), agent session visibility in Issues/Projects, commit-level traceability, and Copilot coding agent management REST APIs. JetBrains launched Central with governance and execution capabilities, warning the industry was about to "repeat the cloud ROI crisis." Kiro added MCP Registry Governance and Model Governance. Chainguard launched a secure-by-default dependency front door for AI agents. The pattern was unmistakable: enterprises wanted AI coding agents, but on their terms — with audit trails, policy enforcement, model governance, and the ability to slow agents down when reliability demanded it.

• Spec-driven development emerged as the leading methodology for professional AI-assisted coding. The "coding was never the bottleneck" thesis gained empirical support throughout March: Agoda's engineering analysis showed that specification and verification — not code generation — were the actual constraints. Patrick Debois presented four patterns for AI-native development at QCon: transitioning from producer to manager, focusing on intent over implementation, moving from delivery to discovery, and managing agentic knowledge. Simon Willison argued for "compound engineering" — running retrospectives after each agent-assisted project to improve future instructions. A freeCodeCamp tutorial introduced "spec-writer," a Claude Code skill generating structured specs with explicit [ASSUMPTION] tags before any code was written. AWS's Strands Labs shipped AI Functions with an @ai_function decorator for specification-driven programming. The convergence was clear: the developer's primary deliverable was shifting from code to specifications and acceptance criteria, with AI agents handling the translation from intent to implementation. Teams that optimized for specification quality were seeing dramatically better outcomes than those focused solely on code generation speed.