Shaping agentic AI
Agentic AI Foundation
Vstorm was the first agentic AI engineering company to join the Agentic AI Foundation and received its 2026 Open Agentic AI Innovation Award.
Read the announcementAgentic AI tech stack selection
A stack you can leave without a rewrite
Agents that keep working when a model or vendor changes. Framework, models and hosting chosen on tests with your own data, with a written plan to switch any of them.
50+ workflows agentified since 2017
-
~20% → ~40% conversion to a final saleThree-agent product advisor guiding customers through print-order configuration.
Read case study -
42% → 98% raw LLM vs guideline-executing accuracy on the open 50-scenario benchmarkHIPAA-compliant guideline-executing triage — 42% to 98% on the open 50-scenario benchmark.
Read case study -
Up to 1 hr → ~2 min to create a workflowText-to-workflow agents building validated node graphs inside the platform.
Read case study -
Supply chain intelligence agents reducing manual coordination overhead.
-
AI agent implementation for a global automotive enterprise.
-
Every inquiry answered in English and Arabic, from ARIJ's own knowledge baseBilingual English/Arabic agent inside ARIJ's Moodle environment, answering only from ARIJ's own knowledge base.
Why stack choices age badly
Stacks picked from demos, on criteria nobody wrote down.
Lock-in, cost at volume and skills gaps surface a year later, when switching is expensive.
- 01
Where it stallsThe stack is chosen after a convincing demo.
What we do Criteria agreed and weighted before any demo.
- 02
Where it stallsOptions compared on the vendor's own examples.
What we do The shortlist is tested on your data.
- 03
Where it stallsNobody plans how to leave.
What we do Every layer gets a written exit path.
How we choose
Criteria first, then evidence from your data
Your criteria, tested by spikes on your own data.
- 0103
Weighted criteria, then shortlists
Weighted criteria rule options out before anyone falls for a demo.
- Criteria weighted with your team
- Shortlists for all five layers
- Constraints that rule options out
- 0203
Spike tests on your data
Two or three options per layer run on your data, and the measurements decide.
- Accuracy and cost per task
- Latency and glue code measured
- Hosted and open-weight models compared
With Mixam's print-order agent, conversion to a final sale went from about 20% to about 40%. - 0303
A decision you can reverse
The record names each layer's choice, the options rejected and why, and the trade-off you accept.
- A decision record per layer
- An exit path per layer
- Spike code in your repositories
Engineering in the open
We help build what agentic AI runs on
Open standards, open-source tooling and published research, all public, so you can check our engineering before you hire us.
Open source
We build open-source agent tooling that other engineering teams install
10M+ PyPI downloads
4,000+ GitHub stars
33 public repositories
Framework contributors
When the framework breaks, we fix it at the source
We maintain our own open-source agent framework and contribute fixes to the frameworks our agents run on, through official partnerships with their teams. Breaking changes reach us before they reach your codebase.
Research
Our engineers and PhDs publish peer-reviewed research
Peer-reviewed papers on agent architectures, evaluation and AI security, and an open benchmark anyone can rerun against our results.
Track record
Frontier AI teams now build on what we shipped first
Our client agents were in production before most of today's agent frameworks had a stable release. You get a team that already knows what breaks between a promising demo and a system people rely on every day.
Case studies
Agent systems our clients run
Engineering workflows built in minutes, validated before they run
The agent drafts each workflow and checks it in several steps before it runs, so engineers skip the manual setup.
Up to 1 hrto prepare a workflow by hand
~2 minto create one with the agent
A print-order advisor that lifted conversion to a final sale
Three agents and 15 tools walk customers through print-order configuration, across more than a billion product combinations.
~20% → ~40%conversion to a final sale
96%+web-to-print orders handled end to end
Clinical triage that executes the guideline, step by step
The model answers the assessment questions. The guideline decides the disposition, the same way every time.
42% → 98%accuracy, raw LLM vs guideline-executing
96%weighted F1 across 566 scenarios
Engagement models
Ways to start
Each path starts with a free call with a founder or senior AI engineer. A build goes ahead only after its Proof of Value meets the agreed threshold.
Start with one workflow
We map one workflow with your experts and agree a success threshold before any build.
Best whenYou want the success test agreed before you commit to a build.
Layers we compare
The layers a stack decision covers
Typical options per layer. Your criteria pick the shortlist.
-
-
- pgvector
- Qdrant
- Milvus
-
- Logfire
- Langfuse
- OpenTelemetry
-
- Your cloud
- Self-hosted
FAQ
Questions about agentic AI stack selection
What does agentic AI tech stack selection include?
A choice for every layer an agent system depends on: the agent framework, the models, retrieval and the vector store, observability, and where the system is hosted. Each layer is chosen against criteria you agree up front, such as lock-in, cost per task, your team's skills, compliance and data residency, and shortlisted options are spike-tested on your data. You get a written decision per layer with the trade-offs and rejected options, an exit path for every layer, and the spike code in your repositories.
You contribute to Pydantic AI. Will you just recommend it?
No. We recommend per use case. Our engineers contribute to Pydantic AI and Vstorm is an official Pydantic implementation partner, so we know where it fits and where it does not. LangGraph, LlamaIndex or a smaller set of plain libraries can be the better choice for a given use case, and the spike results and the decision record show the reasoning, so you can check it.
Hosted or open-weight models: how do you decide?
On your use case and your constraints. Hosted models are quicker to start with and need little operations work. Open-weight models can run in your own environment, which matters for data residency, regulated data and cost at high volume, but your team has to run them. We test the shortlisted models on your tasks and keep the model behind an interface, so changing it later is planned work rather than a rewrite.
What is a spike test?
A short, throwaway build that runs a shortlisted option on a slice of your real case: your documents, your tools, your edge cases. It measures what the criteria ask for, such as accuracy on your tasks, cost per task, latency and how much glue code the option needs. Spikes replace vendor benchmarks with evidence from your own data.
What is an exit path, and why write it down now?
For each layer, the exit path says what you would switch to, what the switch would touch and roughly how much work it is. Writing it down while the decision is fresh keeps the option open: interfaces are drawn so the model, the vector store or the tracing tool can be replaced without rebuilding the agent. A vector store or embedding model change also means re-indexing, and the exit path prices that in.
We already have a stack. Can you review it instead?
Yes. We run the same criteria against what you have, spike the alternatives only where a layer looks weak, and tell you which parts to keep. If an agent built on it has stalled, an AI agent audit is usually the better starting point.
Talk to the engineers
Bring the stack question you are stuck on
Bring your constraints and options; we say what we would test first, and why.
From the blog
Reading on frameworks, models and tooling
Agentic AI
Single agent vs multi-agent systems: when one agent is not enough
One agent or several? Deterministic orchestration beats both when the flow is predictable. Pros, cons and the architectures, with production examples.
Engineering
Is higher accuracy always the goal?
A “minimum 95% accuracy” promise can be worth nothing. An interactive walkthrough of precision, recall and what to ask before any number goes in writing.
Engineering
Agentic RAG: Architecture, Knowledge, Access Control
How agentic RAG changes architecture, knowledge and access control: corrective vs self-reflective designs, what they cost, and where permission checks belong.