First agent in production. You can replay a bad run, not argue about it.
AI Agent Hub
LangGraph holds the graph. LangSmith shows the trace. Braintrust blocks a prompt change that quietly got worse. Modal runs the jobs. Relevance is there so ops can build helpers without opening a ticket.
- Layers 01
5
- Average score 02
83/100
- Free tier 03
5of 5
- Budget 04
$300 to $1,200 / mo
- Team 05
1 to 3 engineers, one owner
[ 01 ] Layer by layer
The tool, and why it is in this set.
- Layer 01
Orchestration
LangGraphLangGraph
AI AgentsOpen sourceFrom $0Explicit state and durable checkpoints. A failed run can be replayed instead of guessed at.
82/100 - Layer 02
Observability
LangSmithLangSmith
AI InfraFreemiumFrom $0Every model and tool call is a trace. A production miss becomes a regression test in one step.
84/100 - Layer 03
Evaluation
82/100 - Layer 04
Compute
ModalModal
AI InfraUsage-basedFrom $0 + computeTool workloads and batch jobs on demand. No cluster to babysit.
85/100 - Layer 05
Ops surface
Relevance AIRelevance AI
AI AgentsFreemiumFrom $0Supporting agents for the business team, without adding work to engineering.
82/100
[ 02 ] Other stacks
Different job, different set.
Automation Stack
n8n for the volume. Zapier only for the connector nobody else has.
B2B SaaS starter
Under 15 people. Free tiers that actually work.
Ops, no engineers
Nothing here needs a repo, a deploy, or someone's afternoon in GitHub.
Revenue signal
Clay in front of the CRM. Cap credits and domains in week one.
Lean AI support
The agent takes first contact. Plain, Notion, PostHog and Linear do the rest.