In this week’s beta of enterprise software strategy, B Capital’s Daisy Cai argues that the rise of AI agents is not a terminal threat to traditional software development but rather a catalyst for a fundamental re-architecture she terms “tech restacking” — a shift where foundational layers like infrastructure, middleware, and application logic are being rebuilt around agentic workflows, with significant implications for cloud economics, open-source sustainability, and enterprise IT agility as of mid-2026.
The Restacking Thesis: Beyond AI Hype to Structural Change
Cai’s thesis centers on the observation that while generative AI captured headlines in 2023-2024, the real transformation now underway is less about chatbots and more about how AI agents are forcing a reevaluation of the software stack’s dependencies. Unlike earlier waves of automation that merely replaced repetitive coding tasks, today’s agentic systems — powered by large language models (LLMs) fine-tuned for tool use and reasoning — are beginning to orchestrate complex, multi-step workflows across services, effectively becoming a latest middleware layer. This isn’t theoretical: early adopters are reporting 30-40% reductions in time-to-deploy for internal developer platforms when agents handle environment provisioning, dependency resolution, and even basic bug triage. The shift echoes the move from monoliths to microservices a decade ago, but with agents acting as intelligent glue rather than just APIs.
Under the Hood: How Agents Are Rewriting the Stack
At the infrastructure layer, we’re seeing cloud providers expose new primitives specifically for agent orchestration. AWS’s Bedrock Agents now support stateful workflow execution with built-in retry logic and audit trails, while Google’s Vertex AI Agent Builder integrates directly with Terraform for infrastructure-as-code generation. On the application side, frameworks like LangChain and LlamaIndex are evolving beyond simple retrieval-augmented generation (RAG) to support dynamic agent chaining — where one agent’s output becomes another’s input, with built-in fallback mechanisms. Crucially, these aren’t just LLM wrappers; they’re leveraging new CPU instructions like Intel’s AMX and AMD’s XDNA2 for efficient transformer inference at the edge, reducing latency for agent decision loops from seconds to sub-500ms in optimized cases. This hardware-software co-design is enabling real-time agent intervention in DevOps pipelines — something impossible with earlier prompt-chaining approaches.
Ecosystem Bridging: Open Source, Lock-In, and the Developer Divide
The restacking trend is already reshaping platform dynamics. Cloud vendors are racing to position their agent frameworks as the central nervous system of enterprise workloads, raising concerns about new forms of lock-in. As one anonymous senior engineer at a Fortune 500 financial services firm place it:
We’re trading vendor lock-in at the VM level for lock-in at the agent orchestration layer. If your agents are tightly coupled to a cloud’s proprietary workflow engine, migrating becomes exponentially harder than lifting and shifting VMs.
Conversely, open-source projects like Haystack and Semantic Kernel are gaining traction by offering vendor-agnostic agent abstractions. The Linux Foundation’s newly launched LF AI & Data Agent Interoperability Initiative aims to define standard interfaces for agent-to-agent communication, drawing parallels to how ODF broke document format silos. Early benchmarks demonstrate that interoperable agent systems incur only a 5-8% performance penalty over proprietary equivalents — a trade-off many enterprises are willing to make to avoid future migration cliffs.
Expert Voices: From Skepticism to Strategic Adoption
While vendor narratives paint agent-driven restacking as inevitable, practitioners offer a more nuanced view. Sarah Chen, CTO of a mid-sized SaaS provider specializing in healthcare data, shared her experience:
We piloted agents for automating compliance report generation. The time savings were real — 70% faster drafting — but the hallucination rate in regulatory citations forced us to keep humans in the loop. Agents are juniors, not replacements; they excel at structuring work but still necessitate senior oversight for judgment calls.
This sentiment aligns with field data from Gartner’s 2025 Q4 survey, which found that while 68% of enterprises are experimenting with agent-augmented development, only 22% have fully automated any critical workflow without human validation gates. The most successful implementations treat agents as force multipliers for junior engineers rather than wholesale replacements — a distinction that determines whether restacking leads to productivity gains or technical debt.
What This Means for the Future of Software
The true significance of tech restacking lies not in replacing developers but in redefining what software engineering means. As agents capture over routine integration, environment management, and boilerplate generation, the value shifts toward higher-order skills: prompt engineering, agent behavior auditing, and workflow ontology design. This mirrors the historical shift from assembly language to high-level languages — not a deskilling, but a reskilling. For enterprises, the immediate priority is evaluating whether their current toolchain supports agent interoperability and observable workflows. For developers, it’s about learning to design systems where agents are first-class participants, not just tools. And for the industry at large, restacking offers a path forward where AI doesn’t just automate the past but helps build a more adaptable, intelligent software foundation — one agent orchestration at a time.
