THINXSTER
Blog/AI Agents
AI Agents10 min readAugust 9, 2026

What Tools to Build AI Agents With: A Stack Map for People Who Ship

The agent tooling landscape is six layers deep and mostly noise. Here's what each layer does, which tools are worth learning, and what to skip.

RK
Ryan Korsz
Founder & CEO, Thinxster

TL;DR

The agent tooling landscape is six layers deep and mostly noise. Here's what each layer does, which tools are worth learning, and what to skip.

→ See how this applies to your business (free 30-min call)

The honest state of agent tooling: there are roughly 200 products calling themselves "agent platforms," about 15 that matter, and 6 distinct layers you need to make a decision about. Most people evaluating this are comparing tools that don't compete with each other.

Here's the stack, layer by layer, with what each one is actually for and what we'd pick.

Layer 1: The Model

Everything else is downstream of this choice, and it's the one people spend the least time on.

  • Claude (Anthropic) — currently the strongest for tool use, long-context reasoning, and following complex instructions without drifting. Claude 5 family for anything where the agent has to hold a multi-step plan. Haiku 4.5 for high-volume, latency-sensitive classification and routing.
  • GPT family (OpenAI) — broad ecosystem, strong general capability, deep third-party integration.
  • Gemini (Google) — competitive context windows and tight integration with Google's data surface.
  • Open weights (Llama, Mistral, Qwen) — worth it when you have a data-residency requirement, a genuinely high-volume narrow task, or a cost floor that hosted models can't hit. Not worth it for a general agent, because you'll spend the savings on engineering.
  • The practical guidance most teams get wrong: use a strong model for the reasoning and a cheap model for the routing. An agent that calls a frontier model for every trivial classification step is burning 10 to 30x more than it needs to. Tiering is the single biggest cost lever in production agents.

    Layer 2: Orchestration

    This is where the agent's control flow lives — what it does, in what order, and how it recovers.

    Code frameworks:

  • The model provider's own SDK plus plain code — genuinely underrated. Most production agents are a loop, a tool registry, and error handling. You don't need a framework for that, and frameworks add debugging surface.
  • LangGraph — explicit state machines for agents. The right answer when your agent has real branching and you need to reason about states.
  • CrewAI, AutoGen — multi-agent orchestration. Impressive demos, harder to make reliable. Use when you genuinely need role separation, not because it sounds sophisticated.
  • Agent SDKs from the model providers — increasingly the default starting point, because they track the model's actual capabilities closest.
  • No-code and low-code:

  • n8n — self-hostable, node-based, with real branching and error handling. Our default for business automation with AI steps. Best value in the category.
  • Make — friendlier visual builder, weaker at complex logic, costs escalate with operation volume.
  • Zapier — easiest to start, most expensive at scale, thinnest logic. Fine for a two-step integration, wrong for an agent.
  • The rule of thumb: if the workflow has more than about a dozen steps or any real error handling, no-code becomes harder than code, not easier. The visual canvas that made step 5 easy makes step 40 unreadable.

    A framework doesn't make an agent reliable. Constrained scope, good tools, and real evals do.

    Layer 3: Tools and Actions

    An agent is only as useful as what it can actually do. This layer is the difference between a chatbot and an agent.

  • MCP (Model Context Protocol) — the emerging standard for connecting models to tools and data sources. If you're building tool integrations today, build them as MCP servers. The ecosystem consolidation is real and moving fast.
  • Function calling — the native mechanism. Well-described, narrowly scoped functions outperform broad ones dramatically. An agent given twelve precise tools reliably beats one given three vague ones.
  • Browser automation — Playwright and the hosted equivalents. Powerful, slow, brittle. Use it when there's no API, not as a first choice.
  • Code execution sandboxes — for anything involving calculation, data transformation, or file manipulation. Letting the model write and run code beats asking it to compute in its head, every time.
  • Layer 4: Memory and Retrieval

  • Vector databases — Pinecone, Weaviate, Qdrant, pgvector. The honest take: pgvector inside Postgres you already run is the right answer far more often than the market implies. Dedicated vector infrastructure earns its cost above roughly ten million vectors.
  • Hybrid search — combining keyword and semantic search consistently beats pure vector retrieval on real corpora. Pure semantic search fails on exact identifiers, part numbers, and names.
  • Structured memory — for most business agents, the "memory" that matters is a CRM record, not an embedding. Don't build a vector store to remember a customer's phone number.
  • Retrieval quality is where most agent projects actually fail, and it's almost always a chunking and ranking problem rather than a model problem.

    Layer 5: Voice

    If the agent talks to humans on a phone, this is its own stack:

  • Vapi, Retell, Bland — infrastructure for real-time voice agents. They handle the hard parts: turn-taking, interruption handling, latency budget, telephony.
  • Deepgram, ElevenLabs, Cartesia — the transcription and speech components underneath.
  • The thing nobody tells you: latency is the entire product. Above roughly 800 milliseconds of response delay, humans start talking over the agent and the conversation collapses. Most of the engineering in a good voice agent is shaving milliseconds, not improving the script.

    This is the layer we spend the most time in. We build AI caller agents that reach every inbound lead within 90 seconds of arrival, hold a real qualifying conversation, and book qualified leads straight onto a calendar, with everything written back to a GoHighLevel pipeline.

    90s
    response time on every inbound lead, day or night
    62%
    average lead qualification rate across client accounts

    Layer 6: Evaluation and Observability

    The layer everyone skips, and the reason most agent projects stall at the demo.

  • LangSmith, Braintrust, Langfuse — tracing, evaluation datasets, regression testing for prompts and agents.
  • Your own eval set — 50 to 200 real cases with known-good outputs. This is worth more than any tool. Build it from actual production transcripts.
  • Structured logging of every tool call — you cannot debug an agent by reading its final output.
  • Without evals, every prompt change is a coin flip. You'll fix one failure mode and silently break two others, and you won't find out for a week.

    What We'd Actually Pick

    For a business automation with AI steps: n8n plus Claude plus your existing database. Ship in days.

    For a customer-facing voice agent: a voice infrastructure provider plus Claude plus a CRM as the system of record. The hard part is the conversation design and the CRM write-back, not the model.

    For a genuinely complex multi-step agent: provider SDK plus plain code plus LangGraph if the state machine warrants it, pgvector for retrieval, Langfuse for tracing, and a real eval set from day one.

    What to skip: multi-agent orchestration until a single agent demonstrably can't do the job. Most "we need a crew of agents" architectures are one well-scoped agent with better tools.

    The Cost Model You Should Build Before You Start

    Almost nobody models agent economics before building, and then gets surprised by the first month's inference bill. The model is simple:

  • Tokens per interaction. Include the system prompt, the tool schemas, the retrieved context, and every turn of conversation history. This is usually 5 to 20x larger than people estimate, because the full context gets re-sent on every turn.
  • Interactions per month.
  • Cost per token for the model you chose, split input and output. Input is typically far cheaper than output — design accordingly.
  • Two levers move this number dramatically. Prompt caching on a large stable system prompt can cut input cost by most of its value on repeated calls. Model tiering — routing the trivial classification steps to a small fast model and reserving the frontier model for actual reasoning — routinely cuts total spend by 60 to 80% with no measurable quality loss.

    Run this model at 10x your expected volume before you commit to an architecture. An agent that's profitable at 500 conversations a month and underwater at 5,000 is a design problem you want to find on a spreadsheet rather than on an invoice.

    The Decision That Matters More Than Tooling

    Scope. The agent projects that succeed do one bounded job with clear success criteria and a defined handoff to a human. The ones that fail try to be a general assistant.

    "Qualify inbound leads and book the good ones" ships and produces measurable revenue. "An AI employee for our business" produces a demo, three months of drift, and a quiet cancellation.

    $102M+
    tracked client revenue generated through this stack

    Pick the narrowest valuable job, build the boring plumbing around it, measure it against a real eval set, and expand only once it's holding.

    If you want an agent that produces revenue rather than a demo, [book a free strategy call](/book) and we'll scope the one job worth automating first.

    Free Weekly Briefing

    One AI Marketing Tactic.
    Every Tuesday. Free.

    What's actually working across our client accounts right now — ROAS moves, follow-up sequences, creative angles. The stuff that isn't in any blog post yet.

    No spam. Unsubscribe anytime. 1,200+ business owners already in.

    Ready to Deploy

    SEE THIS IN
    YOUR BUSINESS.

    30 minutes. We scope the exact systems that apply to your situation and give you a plan.

    ★★★★★ Trusted by 47+ local service businesses

    BOOK A STRATEGY CALL →