What Even Is 'Agent Infrastructure'? A Field Guide for 2026
Frameworks = agent loops (LangGraph, Letta, Mastra, CrewAI, Pydantic AI). Runtimes = where agents execute (Cloud Run, your laptop, Letta Cloud). Infrastructure = email, memory, ingress, budgets, audit (AgentMail, Mem0, Hookdeck, Ujex). Observability = traces, evals (Langfuse, LangSmith, Helicone). Tool registries = MCP servers + the like.
The terms swirl. Here's the cleanup, with examples.
Agent frameworks (orchestration)
Code that runs an agent loop: parse, plan, tool-call, summarize, repeat. Examples:
- LangGraph — state machine + checkpointer + streaming
- Letta — full runtime with three-tier memory
- Mastra — TypeScript-first, opinionated structure
- CrewAI — multi-agent, role-based
- Pydantic AI — type-safe, function-calling-first
- OpenAI Agents SDK — first-party, OpenAI-flavored
Agent runtimes (execution)
Where the agent code physically runs. Examples:
- Your laptop, your server, your Cloud Run / fly.io / Lambda
- Letta Cloud — managed runtime that hosts Letta agents
- Coding-agent runtimes (Claude Code, Cursor, Aider) — agent + IDE bundled
Agent infrastructure (substrate)
The things every agent needs that aren't the loop: email, memory, ingress, budgets, mobile approval, audit. Examples:
- AgentMail / Nylas / Composio — email
- Mem0 / Letta / Zep / Ujex Recall — memory
- Hookdeck / Cloudflare Tunnel / Ujex Ingress — webhooks
- Stripe / Ujex Governor — budgets
- Twilio / FCM / Ujex Mobile — approval channels
- OpenLIT / hash-chain libs / Ujex Audit — audit
Agent observability (introspection)
Watching what the agent does at runtime. Examples:
- Langfuse — open-source traces, evals, prompts
- LangSmith — LangChain's hosted observability
- Helicone — proxy-shaped observability
- Honeycomb / DataDog — generic APM, agent-friendly
Agent evals (testing)
Measuring agent quality. Often overlapping with observability. Examples:
- Promptfoo, Langfuse evals, Phoenix, Inspect
Agent tool registries (capability)
Plug-in tools the agent can use. Examples:
- MCP servers — the standard (AgentMail, Nylas, Composio, Ujex Postbox MCP)
- Tool registries (vendor) — Composio, OpenAI's tool registry, Ujex Tools
Where each tool sits in your stack
┌──────────────────────────────────────────────────────┐
│ Agent observability (Langfuse / LangSmith) │
├──────────────────────────────────────────────────────┤
│ Agent framework (LangGraph / Letta / Mastra) │
├─────────┬──────────────────┬────────────┬────────────┤
│ Memory │ Email │ Ingress │ Audit │
│ Mem0 │ AgentMail │ Hookdeck │ Hash-chain │
│ Letta │ Nylas │ CF Tunnel │ libs │
│ Recall │ Postbox │ Ingress │ Ujex Audit │
├─────────┴──────────────────┴────────────┴────────────┤
│ Runtime (Cloud Run / your laptop / Letta Cloud) │
├──────────────────────────────────────────────────────┤
│ Models (Anthropic / OpenAI / Gemini / open) │
└──────────────────────────────────────────────────────┘Where Ujex fits
Ujex is in the "Memory + Email + Ingress + Audit" row, plus Mobile and Governor. We're explicitly not a framework, runtime, or model vendor. Compose with whatever framework you already use.
Where Celistra fits
Celistra is in the "Runtime" row — but for processes you supervise from a fleet, not for a single hosted agent. It's a different shape from Letta Cloud or Cloud Run; closer to "your laptop, but federated."
Reading list
FAQ
Where do I start?
Pick a framework you like (LangGraph for state-machine flexibility; Letta if you want batteries included). Then pick infra: try the free tiers of AgentMail / Mem0 / Cloudflare Tunnel and see what sticks. When you outgrow vendor sprawl, consolidate on Ujex or self-host.
Is this map definitive?
No — the boundaries shift quarterly. Letta now has a hosted runtime (boundary blurs with frameworks). AgentMail does some PI scoring (boundary blurs with audit/security).