Every team building agents makes an implicit bet they rarely price into the decision: the framework they pick today will still be the right one in twelve months. The data says otherwise.
A 2026 enterprise survey across IT leaders found that 57% had spent over $1 million on AI platform migrations in the previous year alone. The average migration project lost $315,000. And the problem compounds — agentic workflows increase switching costs compared to simple chat completions, because prompt patterns, tool definitions, and guardrails are all framework-shaped. Pick wrong, and the bill shows up twice: once to build, once to leave.
Breaking-change cycles are the rule
LangChain shipped three incompatible breaking-change cycles in under three years (v0.1 → v0.2 → v0.3 → v1.0). Each required coordinated upgrades across langchain-core, langgraph, and a dozen integration packages simultaneously. One developer's post-mortem after six months of production issues described a with_structured_output() bug that silently dropped tool config — the model hallucinated tool calls and returned structured JSON with no error signal. The pipeline had been degrading for weeks before someone traced it to the framework layer.
AutoGen, put into maintenance mode by Microsoft in early 2026, stranded organizations that had built multi-agent systems on its specific orchestration model. Anthropic's Claude Agent SDK was renamed in late 2025, carrying breaking changes governed by Anthropic's Commercial Terms of Service — even when used to power customer products.
Frameworks that ship fast and control their API surface can break your agents without breaking their own tests.
Agentic workflows make switching harder
Andreessen Horowitz's 2025 Enterprise AI Report captured the paradox: "The rise of agentic workflows has started making it more difficult to switch between models. As companies invest in building guardrails and prompting for agentic workflows, they're more hesitant to switch to other models."
This is the lock-in flywheel. You invest in prompts, tool schemas, guardrails, and evaluation suites for one framework. When a better model or framework arrives, the switching cost is the entire orchestration layer you built on top, not just the API call pattern.
Example: when OpenAI raised enterprise pricing 20–40% in 2024 and retired GPT-4 base in 2025, teams had weeks to re-evaluate and re-prompt for GPT-4o. Those on vendor-agnostic infrastructure routed to alternatives. Those deeply coupled to OpenAI's SDK re-architected.
The multi-model maturity spectrum
Most teams that claim multi-model support are not multi-model. A 2026 industry analysis defines four levels:
- Level 0 — Single vendor, no fallback. Acceptable pre-PMF, a risk in production.
- Level 1 — "Marketing multi-model." Production on one provider with an untested fallback config. Most companies that claim multi-model are here.
- Level 2 — Routed multi-model. Different request types to different providers. Real cost optimization, real redundancy.
- Level 3 — Continuously evaluated. Every change re-runs the eval suite on multiple providers; the cheapest that passes quality bar wins.
The gap between Level 1 and Level 2 is where the lock-in cost lives. Level 2 requires a shared abstraction layer for models, tools, and agent definitions — exactly what most frameworks do not provide out of the box.
What portable agent infrastructure requires
Decoupling from any single vendor or framework needs three things:
- LLM abstraction — a way to call any provider through the same interface, with per-request routing, failover, and budget enforcement. LiteLLM and OpenRouter exist for this, but they must be in the architecture from day one, not bolted on after migration one.
- Open standards for tools and context — MCP (Model Context Protocol) is now adopted by every major framework and provider. It decouples tool definitions from specific runtimes. An agent's knowledge and memory should be plain portable formats (Markdown, JSON) that survive a platform change.
- Framework-agnostic deployment — the infrastructure should not dictate which framework your agents run on. Currently, LangGraph Platform runs LangGraph agents; CrewAI Factory runs CrewAI agents. That is the lock-in chain at the infrastructure level.
Alpi is built around all three. It uses LiteLLM at its core to support 100+ providers through a single OpenAI-compatible interface, with per-agent model config and daily budget enforcement. Knowledge and memory are plain Markdown and JSON — you own them, not the platform. ALP (Alpi Link Protocol) is Apache-2.0 from day one, with no registry, no discovery service, and no telemetry. Your agent definitions are not locked to Alpi's runtime; portable documents outlive any one execution layer.
The only durable strategy
OpenAI dropped from 50% to 27% market share between 2023 and 2025 while Anthropic climbed from 12% to 40%. DeepSeek V4 Flash delivers comparable quality at 90% lower inference cost than GPT-4o. The 10–100× price spread between frontier models and open-weight alternatives means single-provider strategies leave savings and resilience on the table.
Gartner predicts that by 2028, 70% of organizations building multi-LLM applications will use AI gateway capabilities, up from under 5% today. The question is whether you build portability in now, or pay the $315K migration later.
uv tool install alpi-agent