The Frameworks
The 2026 mainstream OSS lineup:
- LangGraph (LangChain Inc.): stateful, graph-based agent orchestration. Explicit state machine, persistent memory via checkpointers (SQLite, Postgres, Redis), and human-in-the-loop interrupts. Strongest fit when you need deterministic control over the agent’s workflow.
- CrewAI (CrewAI Inc.): multi-agent crews with role-based collaboration. Each agent has a role, goal, backstory, and tool list; the crew orchestrator assigns tasks. Fast to prototype; opinionated about structure.
- AutoGen (Microsoft Research, v0.4 GA Jan 2025): event-driven multi-agent conversations. Asynchronous, message-passing model. Best for research-style agentic workflows and complex agent-to-agent reasoning.
- PydanticAI (Pydantic team): type-safe, dependency-injection style agents. Lighter framework, strong on validation and Python idioms.
- Mastra (TypeScript): node-friendly agent framework with built-in evals and workflows; growing fast among full-stack teams.
- Anthropic Agent SDK and OpenAI Agents SDK: vendor-blessed but mostly OSS, narrower scope, integrate cleanly with each vendor’s tool ecosystem.
Mental models differ enough that switching mid-project is painful — pick deliberately.
When OSS Fits
Custom agent architectures beyond what vendor platforms offer (multi-step reasoning chains, novel orchestration patterns, research workflows). Cost-sensitive high-volume deployments where platform fees ($2/conversation list price for Agentforce) exceed token cost by 5–10x. Teams with strong Python or TypeScript engineering capacity and existing ops maturity (CI/CD, observability, on-call). Air-gapped or sovereign deployments where the vendor SaaS surface is a non-starter. Workflows that span systems no platform integrates well (legacy mainframes, regional banking cores, custom industry stacks).
When Platforms Win
Rapid deployment in 6–12 weeks with native CRM data access. Enterprise governance out of the box: audit trails, RBAC, content moderation, regional data residency. Salesforce/HubSpot/Microsoft ecosystem integration where the agent needs to call dozens of platform features (Flows, journeys, custom objects). Built-in safety and brand controls equivalent to Salesforce Trust Layer. Vendor-managed model upgrades and prompt regression testing.
For most line-of-business CRM use cases, platforms beat OSS on time-to-value and total cost of ownership through year two. After year three, the math depends on volume and customization depth.
Hybrid Approach (Most Common in 2026)
Use platform agents for customer-facing surfaces where compliance, integration depth, and brand-safety matter most. Use OSS frameworks for internal experimentation, specialized analytical workflows, R&D, and high-volume backend pipelines (lead enrichment, deduplication, churn modeling). The two coexist; the question is which workload sits where, not which framework “wins.”
A common reference architecture: Agentforce or Breeze owns the customer conversation; LangGraph workers handle async enrichment and analytics; both write to the same Data Cloud / Data Lake; observability via shared OpenTelemetry pipeline.
Decision Criteria
- Latency-critical, customer-facing → platform.
- Compliance-heavy (regulated industry, EU operations) → platform with documented controls.
- High-volume backend → OSS for unit economics.
- Novel orchestration not yet supported → OSS.
- Small team, 12-week deadline → platform.
Common Failure Modes
- Building a “framework” on top of OSS that nobody else on the team understands.
- Choosing OSS for cost reasons then spending more on infra and on-call than the platform fee.
- Choosing the platform then fighting it for a use case it wasn’t built for.
- Skipping evaluation infrastructure either way — it’s mandatory regardless of framework.
What to Do This Week
Inventory your in-flight agent projects. For each, write the one-sentence answer to “why this framework, not the other one.” If the answer is “because that’s what we used last time,” reopen the decision.
