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 announcementAI agent performance optimization
Cheaper and faster agents, with accuracy held to a baseline
Your agent reaches its cost and latency target without losing accuracy. One change at a time, each rerun against a baseline built from your own cases.
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 -
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.
-
Supply chain intelligence agents reducing manual coordination overhead.
-
AI agent implementation for a global automotive enterprise.
Why tuning goes in circles
Without a baseline, nobody knows what a fix broke.
Trimmed prompts can lose accuracy, bigger models slow runs, and only a rerun shows it.
- 01
Where it stallsA bigger model for every step.
What we do Each step gets the smallest model that passes.
- 02
Where it stallsFive changes shipped in one go.
What we do One change at a time, each one retested.
- 03
Where it stallsGuessing where time and money go.
What we do Every step priced from traces first.
How we tune an agent
Profile, tune, hand over
Every before-and-after is measured on the same set of your cases.
- 0103
Profiled from real runs
We trace real or recorded runs and price each step: model, tokens, tool calls, waits.
- Cost and latency per step
- Baseline eval on your cases
- Costliest steps ranked first
- 0203
Changed one thing at a time
Cheapest changes go first; any change that lowers the baseline score is rolled back.
- Prompts and context trimmed
- Caching and per-step routing
- Retrieval and tool calls tuned
On the open 50-scenario benchmark, guideline-executing triage scored 98%, against 42% for a raw LLM. - 0303
Your team keeps tuning
Everything lives in your repo with a runbook, so your team tests the next cheaper model.
- Eval set and routing rules
- Cost and latency dashboards
- Runbook for testing models
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
AI agents 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. Changes ship only once they meet the agreed target.
Start with the agent you run
We measure your agent on your own cases and agree the target before changing it.
Best whenYour agent runs, but accuracy, cost or speed falls short.
Technology
What we tune with
The model is chosen per step and stays replaceable.
FAQ
Agent performance optimization, answered
What does AI agent performance optimization include?
Profiling from traces, a baseline eval set built from your cases, and changes made one at a time: trimmed prompts and context, caching, model routing per step, batching and fewer tool round trips, plain code in place of steps that never needed a model, and retrieval tuning through chunking, hybrid search and reranking. Each change is rerun on the same eval set. The eval set, routing rules and cost dashboards stay in your repository.
When does an agent need optimization rather than a rebuild?
When its answers are already useful and the problem is the bill, the wait or the share of answers that miss. If the answers are wrong because the process or the data behind them is wrong, tuning will not fix that, and the profile says so before any work starts.
Will cutting cost lower accuracy?
It can, so every change runs against the baseline eval set; a change that lowers the score is rolled back. The exception is a trade-off you decide to keep, with the numbers on both sides in front of you.
Do we have to switch models or providers?
No. Routing sends each step to the smallest model that passes its cases, and that may be the model you use today. The system stays model-agnostic, so hosted and open-weight models can be compared on the same eval set.
What do you need from us to start?
Access to the agent's code, traces or logs from real runs, and a sample of cases with the answers you expect. If the agent has no tracing yet, we add it first and record a sample.
How do you measure latency and cost?
From traces: duration, token counts and model for every step, plus the tool calls each step makes. Cost per run is computed from those token counts, split into cached, uncached and reasoning tokens, at your current provider prices, so before and after are priced the same way.
Does it work with any agent framework?
Yes. Profiling, caching, routing and retrieval tuning apply to agents built with LangChain, LangGraph, LlamaIndex, Pydantic AI or plain code. We tune the framework you already run.
Can our team keep tuning after the engagement?
Yes. The eval set, the routing rules and the dashboards live in your repository, with a runbook for testing a new model or prompt against the baseline.
Start with the traces
See where your agent's time and money go
Book a call, and we go through a sample of your runs together.
From the blog
Reading on context, models and tools
Engineering
Why RAG is not dead: a case for context engineering over massive context windows
Retrieval-Augmented Generation (RAG) and context windows enhance AI accuracy—discover how to optimize your AI workflows effectively
Engineering
Beyond Frontier Models: Testing Lightweight LLMs for Document Processing in RAG
Why bother with small models when capable LLMs are available? Because businesses may face stricter requirements than inidividual users
Agentic AI
Dynamic Tool Discovery: A Systematic Evaluation of 8 Search Strategies for AI Agents
51 tools cost 18,492 tokens per query. Eight discovery strategies — regex, BM25, RAG and more — evaluated on accuracy, latency and token cost.