Skip to content
StackCook
Subscribe

Build a customer-support agent

Answer customer questions from your docs/tickets and hand off to a human when unsure.

Early ranking, based on docs and benchmarks, not yet hands-on tested.

Difficulty
Medium
Status
Early
Last reviewed
2026-10-17

Stack

Framework
  1. #1LangGraph7.6Compare
    Why this score for LangGraph(4 notes)
    Fit for the job 8/10
    Human-in-the-loop interrupts + used in production support build.
    Developer experience 6/10
    Lower-level graph API; more code. Editorial read of docs; no own test yet.
    Reliability & production readiness 8/10
    Checkpointing; production case study (vendor blog).
    Openness 8/10
    MIT.
  2. #2OpenAI Agents SDK7.2Compare
    Why this score for OpenAI Agents SDK(4 notes)
    Fit for the job 8/10
    Handoffs + guardrails map directly to triage/escalation; official CS demo.
    Developer experience 7/10
    Editorial read of docs; no own test yet.
    Reliability & production readiness 6/10
    Demo repo, no production case study found yet.
    Openness 7/10
    MIT; provider portability not verified.
  3. #3Vercel AI SDK7.0Compare
    Why this score for Vercel AI SDK(4 notes)
    Fit for the job 6/10
    General toolkit; handoff/HITL must be built yourself (not verified otherwise).
    Developer experience 8/10
    TypeScript-first. Editorial read of docs; no own test yet.
    Reliability & production readiness 6/10
    No support-agent production case found on this pass.
    Openness 7/10
    Provider-agnostic; license shows NOASSERTION on GitHub API.
Weights
  • Fit for the job: 25%
  • Developer experience: 20%
  • Reliability & production readiness: 20%
  • Openness: 15%
  • Ecosystem & integrations: 10%
  • Momentum: 10%
Model
  1. #1Claude Haiku 5.57.4Compare
    Why this score for Claude Haiku 5.5(5 notes)
    Task quality 7/10
    75.3% τ²-Airline, within ~3pp of Sonnet 5.5.
    Cost per typical job run 9/10
    $0.012/task on τ²-Airline; $0.10/$0.50 per MTok.
    Latency / speed 6/10
    1.6 min/task on τ²-Airline (multi-turn task, not per-reply latency).
    Context & modality support 8/10
    1M context.
    Openness / deployability 4/10
    Closed weights; API + Bedrock + Google Cloud.
  2. #2Claude Sonnet 5.57.1Compare
    Why this score for Claude Sonnet 5.5(3 notes)
    Task quality 8/10
    78.0% τ²-Airline.
    Cost per typical job run 6/10
    $0.16/task; $2/$10 per MTok.
    Latency / speed 8/10
    82 s/task, fastest on the board.
  3. #3GPT-6.1 Sol6.0Compare
    Why this score for GPT-6.1 Sol(5 notes)
    Task quality 6/10
    70.0% τ²-Airline.
    Cost per typical job run 7/10
    $0.10/task; $2/$10 per MTok.
    Latency / speed 5/10
    2.4 min/task.
    Context & modality support 8/10
    1.05M context.
    Openness / deployability 3/10
    Closed weights; OpenAI API.
Weights
  • Task quality: 15% — Support turns are short; policy/tool correctness matters more than peak reasoning.
  • Cost per typical job run: 25% — Many conversations; cost per resolved ticket dominates.
  • Latency / speed: 15%
  • Tool-calling & structured-output reliability: 25% — Handoff and policy-following via tools is the core of the job.
  • Context & modality support: 10%
  • Openness / deployability: 10%
Knowledge
  1. #1Supabase (pgvector)tie-break: more verified evidence7.5Compare
    Why this score for Supabase (pgvector)(5 notes)
    Operational simplicity 8/10
    Same Postgres as app data/tickets. Editorial read of docs; no own test yet.
    Cost at MVP scale 8/10
    Free 500 MB; Pro from $25/mo.
    Scale & performance 6/10
    No independent benchmark used.
    Access control & data governance 8/10
    RLS-scoped retrieval documented.
    Openness 8/10
    Open-source pgvector, self-hostable Postgres.
  2. #2Qdranttie-break: more verified evidence7.5Compare
    Why this score for Qdrant(5 notes)
    Operational simplicity 7/10
    Managed or self-host; separate service to run.
    Cost at MVP scale 8/10
    Free 1GB RAM tier forever.
    Scale & performance 7/10
    Vendor claim only.
    Access control & data governance 6/10
    API keys/JWT; per-doc user filtering is DIY.
    Openness 8/10
    Apache-2.0, self-hostable.
  3. #3Pinecone7.0Compare
    Why this score for Pinecone(5 notes)
    Operational simplicity 8/10
    Fully managed.
    Cost at MVP scale 7/10
    Free Starter; $20 Builder; $50 min Standard.
    Scale & performance 7/10
    Vendor claim only.
    Access control & data governance 6/10
    Namespaces partition data; per-user ACL not verified.
    Openness 3/10
    Proprietary managed service (BYOC exists per banner, not verified).
Weights
  • Retrieval quality & features: 25%
  • Operational simplicity: 20%
  • Cost at MVP scale: 20%
  • Scale & performance: 15%
  • Access control & data governance: 10%
  • Openness: 10%
Tools
  1. #1Intercom MCP server7.3Compare
    Why this score for Intercom MCP server(5 notes)
    Coverage of required actions 7/10
    Read conversations/contacts, read+write Help Center; no reply/close actions listed.
    Security model 8/10
    OAuth scoped to existing permissions.
    Reliability / maintenance 7/10
    Official, hosted; not AU.
    Setup effort 8/10
    Remote endpoint, no server to run.
    Openness / official vs community 6/10
    Official but proprietary hosted.
  2. #2Stripe MCP server7.0Compare
    Why this score for Stripe MCP server(4 notes)
    Coverage of required actions 6/10
    Billing/refund actions only — complements a helpdesk server.
    Security model 8/10
    Human confirmation required for refunds/outbound payments.
    Reliability / maintenance 7/10
    Official.
    Openness / official vs community 7/10
    Open-source toolkit repo.
  3. #3Zendesk MCP server (community)4.6Compare
    Why this score for Zendesk MCP server (community)(5 notes)
    Coverage of required actions 6/10
    Broad ticket/user/macro tools per README (community).
    Security model 4/10
    Community server; uses your API creds; not audited.
    Reliability / maintenance 4/10
    Unofficial; official Zendesk MCP server not released as of Aug 2026.
    Setup effort 5/10
    Self-run.
    Openness / official vs community 3/10
    Source is public but has no license, so it isn't open source.
Weights
  • Coverage of required actions: 30%
  • Security model: 25%
  • Reliability / maintenance: 20%
  • Setup effort: 15%
  • Openness / official vs community: 10%
Evals
  1. #1Langfuse8.0Compare
    Why this score for Langfuse(3 notes)
    Integration effort 8/10
    OTel + framework integrations. Editorial read of docs; no own test yet.
    Cost / free tier 8/10
    Hobby free 50k units/mo.
    Openness / self-host 8/10
    Self-hostable.
  2. #2LangSmith7.1Compare
    Why this score for LangSmith(4 notes)
    Tracing depth & debugging 8/10
    Used in Minimal's production build.
    Integration effort 7/10
    Tightest with LangGraph.
    Cost / free tier 7/10
    Free 5k traces; Plus $39/seat.
    Openness / self-host 4/10
    Managed SaaS; self-host enterprise only (not verified).
  3. #3Braintrust6.8Compare
    Why this score for Braintrust(2 notes)
    Cost / free tier 7/10
    Starter $0, 14-day retention; Pro $249/mo.
    Openness / self-host 4/10
    Proprietary; on-prem enterprise only.
Weights
  • Tracing depth & debugging: 25%
  • Eval features: 25%
  • Integration effort: 20%
  • Cost / free tier: 15%
  • Openness / self-host: 15%
Hosting
  1. #1Vercel7.2Compare
    Why this score for Vercel(4 notes)
    Fit for workload 8/10
    Fluid compute suits I/O-bound LLM calls + streaming chat UI.
    Cost at MVP scale 6/10
    A business support agent needs Pro ($20/mo): Hobby is non-commercial only.
    Deploy DX 8/10
    Editorial read of docs; no own test yet.
    Lock-in / portability 6/10
    Next.js portable; platform features less so.
  2. #2Render7.0Compare
    Why this score for Render(4 notes)
    Fit for workload 7/10
    Long-running services/workers suit background ticket jobs.
    Cost at MVP scale 7/10
    $7/mo 512 MB instance; free instances have limitations; team workspaces (Pro) $25/mo + compute.
    Deploy DX 7/10
    Editorial read of docs; no own test yet.
    Lock-in / portability 8/10
    Docker containers portable.
  3. #3Cloudflare Workers6.9Compare
    Why this score for Cloudflare Workers(4 notes)
    Fit for workload 7/10
    Agents SDK for stateful agents; CPU limits per invocation.
    Cost at MVP scale 8/10
    Free 100k req/day; low paid rates.
    Deploy DX 7/10
    Editorial read of docs; no own test yet.
    Lock-in / portability 5/10
    Workers runtime differs from Node.
Weights
  • Fit for workload: 30%
  • Cost at MVP scale: 20%
  • Deploy DX: 20%
  • Reliability / scaling: 15%
  • Lock-in / portability: 15%

Get ranking updates for this stack

When a pick changes or a new build lands, it's in the weekly email.

Subscribe