Notes from the field.
Short essays and longer analysis from the Steinn Labs team, written for the practitioners we work with.
- 012026.07.04Insight · 9 min
The Hidden Costs of Agentic AI: Token Spend, Latency, and Failure Modes
The advertised cost of an agentic AI system is the model API price per token. The real cost is three to ten times higher once you account for multi-step token compounding, retry overhead, evaluation and judge layers, failure handling, infrastructure, and the engineering time to maintain a system whose behaviour changes when model providers push updates. Most teams discover this in production, not in the pilot. Understanding these costs before you build determines whether the business case holds at real volume.
→ - 022026.07.04Insight · 11 min
Tool Use and Function Calling: How AI Agents Actually Take Action
Tool use is the mechanism that turns a language model into an agent. Without tools, a model can only generate text. With tools, it can query databases, call APIs, send emails, run code, and affect real systems. Function calling is the technical implementation: the model receives a list of available functions with their parameters, decides which function to call and with what arguments, and the host application executes the call and returns the result. The model then uses that result to decide its next step. This loop, decide, call, observe, continue, is the core of how every production agentic system works.
→ - 032026.07.04Insight · 9 min
Building Agentic AI for KYC and Onboarding Workflows
An agentic KYC system collects customer documents, runs identity verification, executes sanctions and PEP screening in parallel, identifies gaps and requests missing information, scores risk, and assembles the complete onboarding file for compliance sign-off. For standard retail and SMB customers, this entire sequence runs in under a minute. JPMorgan Chase's agentic KYC system, reported in production in April 2026, compressed a five-day process to under one minute. The architecture that produces these results is not AI embedded in an existing workflow. It is a full reimagination of the workflow around agentic orchestration from the start.
→ - 042026.07.04Insight · 9 min
Agentic AI in Banking Operations: Where It Works Today vs Where It's Still Risky
Agentic AI is running in production in banking today across fraud monitoring, KYC and onboarding, compliance documentation, and internal operations. These use cases share a common profile: bounded scope, measurable success criteria, reversible or low-consequence actions, and a clear human checkpoint before anything irreversible happens. The use cases that are still risky share the opposite profile: open-ended judgment, high-consequence irreversible actions, or regulatory requirements that currently mandate a human decision-maker. Knowing which category your use case falls into is the most important thing you can determine before starting a build.
→ - 052026.07.04Insight · 8 min
Self-Hosted AI Agents vs Cloud APIs: Architecture Tradeoffs
Self-hosted AI agents run models on infrastructure you control: your servers, your cloud tenant, your data never leaving your environment. Cloud API agents send prompts and data to an external provider's infrastructure and receive responses. The right choice depends on four factors: data residency requirements, cost at scale, latency tolerance, and how much control you need over the model itself. For most regulated industries in the UAE, self-hosted is not a preference. It is a compliance requirement.
→ - 062026.07.04Insight · 10 min
Tamper-Evident Audit Trails for AI Agents: A Technical Primer
A tamper-evident audit trail for an AI agent is a log where every record is cryptographically linked to the record before it, so that any modification to a past record breaks the chain and becomes detectable. The mechanism is a hash chain: each log entry includes the hash of the previous entry, making the log append-only and verifiable. For agentic AI systems in regulated environments, this is the difference between a log that shows what happened and a log that proves what happened, and that distinction matters to regulators.
→ - 072026.07.04Insight · 10 min
AI Agent Failure and Human Oversight: Designing Escalation for Production Systems
AI agents fail in ways conventional software does not: silently, confidently, and across multiple steps before anyone notices. Designing for failure means building three things before you build anything else: a classification of which failures the agent can recover from itself, which require a human decision, and which require the system to stop entirely. Everything else in your escalation architecture follows from that classification. Without it, you are designing for the happy path and hoping the real world cooperates.
→ - 082026.07.04Insight · 7 min
Why Most "AI Agents" Are Just Chained Prompts (And How to Tell the Difference)
Most products marketed as "AI agents" in 2026 are prompt chains: a fixed sequence of LLM calls where each output feeds the next input. This is useful software, but it is not an agent. A real agent perceives its environment, decides its own next step based on what it finds, takes action through real tools, and adjusts its plan when something unexpected happens. The test is simple: if the sequence of steps is fixed in advance by a developer, it is automation. If the system decides the sequence itself based on what it observes, it is an agent.
→ - 092026.07.04Insight · 7 min
LangGraph vs AutoGen vs CrewAI: Choosing an Agent Framework
LangGraph is the default choice for production agentic systems that need explicit state control, human-in-the-loop approvals, and serious observability. CrewAI is the fastest path to a working multi-agent prototype, with role-based abstractions that are intuitive to set up but limited in production reliability. AutoGen, now in maintenance mode under Microsoft, remains relevant for conversational multi-agent patterns and teams on .NET or Azure stacks. For most production builds in 2026, the decision comes down to LangGraph for anything stateful and compliance-sensitive, CrewAI for fast validation before committing to a full build.
→ - 102026.07.04Insight · 7 min
Multi-Agent Systems vs Single-Agent AI: When to Use Which
Use a single agent when your task is well-defined, runs in one domain, and can be completed by one model with the right tools. Use a multi-agent system when the task is too large for one context window, requires genuinely different capabilities running in parallel, or needs one agent checking the work of another. Most real production use cases today are single-agent. Multi-agent adds real complexity and should be justified by a real problem, not architectural ambition.
→ - 112026.07.04Insight · 8 min
Agentic AI Consulting in the UAE: What It Actually Costs in 2026
Agentic AI consulting in the UAE costs between AED 30,000 for a discovery engagement and AED 1,000,000 or more for a production-grade multi-agent system in a regulated environment. A realistic end-to-end build for a mid-sized fintech sits between AED 500,000 and AED 800,000 in Year 1, including discovery, pilot, and production build. Ongoing operations cost 25 to 40 percent of the build cost annually. The wide range is driven by data readiness, compliance requirements, deployment architecture, and whether the system is self-hosted or cloud-based.
→ - 122026.07.04Insight · 7 min
How to Evaluate an Agentic AI Vendor: Five Questions to Ask
To evaluate an agentic AI vendor, ask five things: whether they have production deployments (not just pilots), how they handle agent failure and escalation, where your data goes and how compliance is handled, who actually builds the system, and what support looks like after launch. Their answers will tell you more than any proposal document.
→ - 132026.06.30Insight · 7 min
What Is Agentic AI? A Practical Definition for Business Leaders
Agentic AI is an AI system that can take a goal, break it into steps, decide which actions to take, carry out those actions using real tools and systems, check its own results, and keep going until the goal is done or it needs a human to step in. That is the whole definition.
→