Standardize the team on one agent framework, or let each pod pick?
LangGraph, CrewAI, AutoGen, and 'roll-your-own' are all in production at peer companies. Two years in, every pod that picked separately is bearing its own framework debt.
The question
Our engineering pods have been adopting agent frameworks independently — some on LangGraph, some on CrewAI, some rolling their own asyncio orchestrators. Maintenance load is growing and cross-pod hires take weeks to be productive. Do we standardize on a single framework now (and force migrations), let pods keep choosing, or commit to a thin internal abstraction over multiple frameworks?
The premise
- Team
- ~50 engineers, ~10 actively building AI features, single MLOps engineer. AI work pulls from feature-shipping capacity — any new commitment has to trade against the roadmap. Three pods each own their own agent stack; no central platform team for agents.
- Compliance
- SOC2 Type II in scope. EU customer data subjects us to GDPR plus the EU AI Act's August 2026 GPAI-deployer obligations. Each framework's audit-logging story is different; consolidating reduces compliance surface.
- Stack
- 11 production agents split across 3 frameworks: 5 on LangGraph (customer-facing RAG + workflows), 4 on CrewAI (internal ops automation), 2 on a homegrown asyncio orchestrator (highest-throughput data-pipeline use cases). Each pod owns its framework knowledge; cross-pod hires take ~4 weeks to be productive on the framework they didn't ship before.
- Budget
- Monthly AI spend ~$30K with quarterly board visibility. Approvals required for sustained jumps >20%. Cost-per-outcome metrics in place; finance asks for unit economics by use case. Standardization is mostly engineering time (migrations + abstraction code), not external spend.
When does the maintenance burden of multiple frameworks force consolidation?
When senior-engineer headcount drops below 2 per framework (currently borderline on CrewAI — one departure away), OR when a security incident requires patching across all three at once (hasn't happened, would be the forcing function), OR when cross-pod hire onboarding crosses ~6 weeks. We're not at any of these triggers yet, but the trajectory is clear.
Which framework's lock-in profile is most acceptable?
LangGraph: tightest coupling to the LangChain stack but largest ecosystem and the most-adopted production patterns. CrewAI: lighter, more portable, but smaller community and less production-tested at our scale. Homegrown: zero lock-in, full maintenance burden. We're least uncomfortable with LangGraph because the lock-in is to LangChain — widely-used enough that we'd absorb the cost. Custom is the riskiest long-term once the original author rotates off.
Is a thin internal abstraction over multiple frameworks viable?
Theoretically yes; practically no. The maintenance cost of the abstraction layer + the cost of integration tests across 2-3 frameworks exceeds the cost of standardizing on one. We've underestimated this twice before (event-bus, DB-client wrappers — both got rewritten as direct calls within 18 months).
Counsel's position
Deploy an independent orchestration control plane to centralize audit logging across your existing frameworks, and mandate LangGraph exclusively for new customer-facing workflows to satisfy upcoming EU AI Act requirements.
Verdict
The verdict: Standardize on LangGraph for compliance-bound production workflows — LangSmith integration offers full traces for each graph run, supporting the auditability needed for SOC2 and EU AI Act requirements.
Standardize on LangGraph for compliance-bound production workflows
Given your SOC2 and upcoming EU AI Act obligations, prioritize a framework that natively supports human-in-the-loop checkpoints and predictable execution.
Adopt an independent control plane to unify your fragmented agent stack
Instead of forcing migrations across your three engineering pods, implement an interaction layer that provides centralized auditability across LangGraph, CrewAI, and custom asyncio agents.
Read another verdict
- Build agents that own workflows — or workflows that own LLM calls?
- Set our LLM data retention policy now, or wait for an incident to force it?
- Build our own vertical copilot — or buy from a category vendor?
- Roll AI coding tools across the whole engineering org — and how do we measure it?
- Kill every AI pilot that can't show ROI in 90 days?
- Use AI to flatten middle management this year?
- Stand up a FinOps practice for tokens and GPUs now?
- Replace customer support with AI — or avoid the Klarna outcome?