Resume-driven architecture
Choosing stacks, meshes, and agents because they look good on LinkedIn and conference CFPs — not because traffic or constraints demand them.
Resume-driven architecture is when the system diagram optimizes for the next job posting. Microservices for four teams, CQRS for a CRUD form, blockchain for a loyalty points table, multi-agent frameworks for a cron job — each choice has a conference talk and a weak production rationale.
The pattern is rational for individuals and expensive for companies. Engineers are rewarded for novelty in interviews; products are rewarded for boring reliability. When those incentives collide, you get YAML empires, embedding pipelines without evals, and "AI-native" platforms that still cannot reconcile two databases at month-end.
History is merciless: the same org that cargo-culted Hadoop later cargo-culted Kubernetes, then mesh, then agents. Each wave left scar tissue and a platform team that gates deploys. Consolidation arrives years later dressed as rationalization — after the resumes have already been updated.
The antidote is constraint-first design: pick the simplest architecture that fits measured bottlenecks, headcount, and ownership. If the justification starts with "it will look good when we hire," you are not designing a system — you are drafting a LinkedIn summary with an on-call rotation.
Exhibits
Architecture · 2012–now
Microservices Everywhere
Independent deployability as religion. Correct for some; ruinous cargo-cult for most.
$ Teams split 4-person products into 40 services, then hired SREs to babysit the mesh. Many later consolidated — after years of distributed tracing invoices and on-call trauma. See case studies on premature microservices and big-bang cutovers for named failure modes.
Practice · 2010–2020
CQRS everywhere
Command/query separation as default architecture. Right for high-scale event domains; absurd for to-do lists.
$ Teams split read and write models for CRUD apps that never needed separate projections. Double the schemas, double the bugs, double the on-call surface — for a blog admin.
Practice · 2017–2023
Blockchain for CRUD
Distributed consensus for problems that needed a database and an audit log.
$ Enterprises funded "innovation labs" to put loyalty points and supply-chain spreadsheets on ledgers. Most pilots ended as slide decks; the gas fees and audits did not.
Framework · 2023–now
Multi-agent orchestration hype
LangGraph, CrewAI, AutoGPT cosplay — agents delegating to agents for tasks one function call could do.
$ Autonomous agent swarms burned API budgets on coordination loops that a single prompt and a cron job would have handled. Debugging "which agent lied" became the new on-call sport.
Architecture · 2017–2024
Service mesh everywhere
Sidecar proxies for every microservice. Netflix problems with startup headcount.
$ Istio sidecars on three-service startups burned SRE headcount on mTLS ceremonies nobody could debug. Many teams ripped mesh out before the observability dashboard paid rent.
Practice · 2023–now
Embed everything
If it is text, vectorize it. Semantic search cosplay for problems that needed a better filter.
$ Embedding pipelines for tickets, PDFs, Slack, and the cafeteria menu — then nobody measured retrieval quality. Vector bills and reindex jobs became the product.
Architecture · 2019–now
Data mesh
Decentralize data ownership like microservices for warehouses. Real org problem; often a logo refresh for the same broken pipelines.
$ Consultancies sold domain ownership and federated governance while teams still could not agree on a customer ID. Slideware outpaced catalog tooling by years.
Practice · 2024–now
Eval theater
LLM evals as slideware — metrics that look scientific and never gate a deploy.
$ Dashboards of vibe-check scores that never blocked a release. Green charts for leadership; prod still hallucinated.
Framework · 2023–now
LLM orchestration frameworks
Chains, tools, memory, and agents as a framework — LangChain-class glue sold as architecture.
$ Teams adopted chain/agent frameworks before they had a single reliable tool call. Abstraction layers multiplied while prompt quality stayed flat.