The Two Shapes
OpenAI Assistants is a managed agent framework: it handles state, tool use, file context, and threading on OpenAI's side. You give it a system prompt and tools; it does the rest. Custom agents are what you build with raw API calls plus your own orchestration: more code, more control.
The choice isn't about which is “better.” It's about which trade-off you want.
When OpenAI Assistants Win
- You're shipping fast and the agent shape fits your use case.
- You don't want to manage conversation state yourself.
- You're OpenAI-committed and won't swap providers.
- Your tool use is straightforward (lookups, simple API calls).
- You're a small team that values fewer moving parts.
When Custom Agents Win
- You want to swap models or providers without rewriting.
- You need fine control over conversation state (compression, summarization, branching).
- You're running at scale where managed pricing hurts.
- Your tool use is complex (chained calls, conditional flows, multi-agent).
- You need detailed observability and evals.
- • Prototype + small team + speed → Assistants.
- • Production at scale + multi-provider future → Custom.
- • Migrate from Assistants to Custom when scale hurts or lock-in stings.
The Trade-offs
Assistants give you speed at the cost of lock-in and pricing. Custom gives you control at the cost of months of work. Neither is wrong; they're choices about where you spend.
Migration Between Them
If you've built on Assistants and need to migrate, plan 4–8 weeks per agent. The state model is the hardest part — Assistants stores threads on their side, your custom version has to recreate them. Plan a parallel run period.
What “Custom” Means In 2026
Custom doesn't mean from scratch. Frameworks like LangChain, LlamaIndex, Anthropic's own SDK with conversation helpers, and platform-native edge orchestration all reduce custom agent code by 70%+ vs raw HTTP.
The right question isn't “managed or custom?” It's “at what scale and provider commitment level does managed stop paying off?” Most teams answer that wrong by underestimating future scale.
FAQ
What about Anthropic's equivalent? Less managed, more SDK-driven. Custom-leaning by default.
Can we mix? Yes — some agents on Assistants, others custom. Common at mid-size companies.
Open-source frameworks? Many. Production maturity varies. Test before committing.