AI Agent Development Services

StepTo, a 15-20 engineer Belgrade team, builds AI agents that call your APIs, with confirmation steps and audit logs. Plus a sourced MCP and agent-risk guide.

Reviewed by Igor Gazivoda, Founder & CEO of StepTo · Updated

An AI agent project is an engineering project. A language model is given a goal and a set of tools, and the work is deciding which tools it may call, what it may do without asking, how every action is logged and how you measure whether it behaves. StepTo is a team of 15-20 engineers in Belgrade, Serbia, on Central European Time. Its published AI pages describe building custom AI agents that call your APIs, with confirmation steps, audit logs and human review where a mistake would be costly. StepTo's published rate card runs $25-85/hr, and engagements start in 2-4 weeks.

This page takes the agent angle. For the wider AI practice (LLM integration, RAG, workflow automation) see AI automation solutions. For assistants and bots see chatbot development. To add individual engineers see hiring AI developers, and for an engineer embedded in your own systems see forward deployed engineers.

The sections on the Model Context Protocol (MCP) and on agent risks are neutral explainers built from primary sources, which are listed at the end of the page.

What Is an AI Agent, and When Is a Plain Workflow Enough?

Anthropic's guidance on building agents separates workflows, where models and tools are orchestrated along code paths defined in advance, from agents, where the model directs its own process and tool use. It advises choosing the simplest solution that works, and adds that a single model call with retrieval is often enough. Agents add cost and latency, so they earn their place only where model-driven decisions demonstrably improve the outcome.

PatternWhat it isWhere it fits
A single model call with retrievalOne prompt, with relevant documents fetched and added as context.Questions answered from your own content. Often enough, and the cheapest to run.
A workflowThe model and your tools are chained along a path your code defines in advance.Well-defined, repeatable tasks where you want predictable steps and cost.
An agent with approval gatesThe model chooses which tools to call, and a person confirms the actions that matter.Multi-step work where the path cannot be fixed in advance but a wrong action is costly.
A fully autonomous agentThe model directs its own process and tool use with no checkpoint before it acts.Rarely the right first build: errors can compound and every extra permission widens the blast radius.

What Does a StepTo AI Agent Engagement Cover?

Each point below restates what StepTo already publishes on its chatbot development, forward deployed engineer and AI automation pages. Integrations with the systems you run and the choice of models are covered on the chatbot development page.

Tool calling with confirmation and audit logs

StepTo's published AI pages describe agents that call your APIs, with confirmation steps and audit logs, so an action taken on a person's behalf can be reviewed afterwards.

Human review where a mistake is costly

Agentic workflows are described with human review wherever a wrong output would be expensive, and with logging so the behaviour can be audited.

Guardrails and evaluation

StepTo builds assistants and agents wrapped in guardrails, with an evaluation discipline that separates a demo from something you can put in front of customers.

A published example of the human-review pattern

The AI support automation case study describes a 3-person dedicated squad that built a RAG system for a B2B SaaS support team over 4 months. As published there, every AI draft on a ticket routed to a support agent is approved, edited or rejected by that agent before it reaches a customer, 38% of tier-1 tickets were auto-resolved with no human touch, and average first-response time fell 70%. It is a RAG project rather than an agent with write access. It is published as an example of a human review queue and confidence thresholds in a StepTo build.

What Is an MCP Server, and What Does Building One Involve?

The Model Context Protocol specification (version 2025-11-25) describes MCP as an open protocol for connecting LLM applications to external data sources and tools. It uses JSON-RPC 2.0 messages between hosts (the LLM applications), clients (connectors inside a host) and servers (services that provide context and capabilities). A server can offer resources (context and data), prompts (templated messages) and tools (functions for the model to execute).

Building an MCP server for your API therefore means designing a set of tools an agent can safely call. The checklist below applies to any tool layer between an agent and your systems, whether or not it is exposed over MCP.

  • Narrow tools with typed inputs

    Each tool does one job, has a clear description and a typed input schema. Anthropic's agent guidance says tool definitions deserve as much prompt-engineering attention as the prompt itself, and recommends making tools hard to misuse.

  • Authorization in the system behind the tool

    The permission check belongs in the downstream system, not in the model. OWASP's excessive-agency guidance recommends checking every request an extension makes against policy, and limiting permissions to the minimum necessary.

  • The MCP specification says tools represent arbitrary code execution, that hosts must obtain explicit user consent before invoking any tool, and that tool descriptions and annotations should be treated as untrusted unless they come from a trusted server.

  • Evaluation on many example inputs

    Run a wide range of sample inputs against the tools, look at the mistakes the model makes, and iterate on the tool design. Test in a sandbox before production.

  • Logging of every call

    Record which tool was called, with what arguments, for whom, and with what result, so a failure or an abuse attempt can be traced.

StepTo's published AI pages describe agents that call tools and APIs. None of them lists MCP server development as a service, so this section is an explainer and not an offer.

What Goes Wrong With AI Agents, and Which Controls Address It?

Four failure modes recur in the primary sources. The controls are paraphrased from them; read the originals before you design permissions.

Excessive agency

OWASP LLM06:2025 traces damaging actions to three root causes: excessive functionality, excessive permissions and excessive autonomy. Its controls include human approval of high-impact actions, least privilege, and authorization enforced in downstream systems.

Source: OWASP LLM06:2025 Excessive Agency

Prompt injection

OWASP LLM01:2025 describes indirect prompt injection as content from an external source, such as a website or a file, altering the model's behaviour. Its mitigations include least privilege, human-in-the-loop controls for privileged operations, segregating external content and adversarial testing.

Source: OWASP LLM01:2025 Prompt Injection

Compounding errors

Anthropic's guidance on building agents warns that autonomy brings the potential for compounding errors. It recommends extensive testing in sandboxed environments, stopping conditions such as a maximum number of iterations, and pausing for human feedback at checkpoints.

Source: Anthropic, Building effective agents

Trusting a tool server too much

The MCP specification states that it cannot enforce its security principles at the protocol level and that implementors should build consent and authorization flows, access controls and clear documentation of security implications.

Source: MCP specification, 2025-11-25

What Does AI Agent Development Cost at StepTo?

StepTo publishes three engagement options, and its AI automation page describes AI work under each of them. Full terms are on the pricing page.

Staff augmentation

From $4,500 per developer per month

A 3-month minimum applies and billing is weekly. The engineer is placed directly within your team.

Dedicated team of three

From $13,500 per month

Billed monthly. The published AI support case study used a 3-person dedicated squad.

Project or MVP

From $9,000

A defined solution delivered from discovery through production deployment, with a fixed scope and an agreed timeline. $9,000 is the starting point for custom software projects and MVPs generally, not a quote for an agent.

The chatbot development page publishes the ranges for action-taking assistants.

Published rate card

StepTo's published rate card runs $25-85/hr by seniority. Engagements start in 2-4 weeks.

Junior
$25-35/hr
Mid-level
$35-45/hr
Senior
$50-65/hr
Lead
$65-85/hr

To scope a first agent with named roles, use the spec a team tool, or see staff augmentation and dedicated development teams.

How Do Other Vendors Present Their AI Agent Work?

Each cell reports what the vendor's own page says as of 2026-11-20. These vendors win on scale, certifications and partner status, which StepTo does not claim. StepTo publishes a rate card; the pages read for the vendors below did not show one.

VendorAgent offer, as reportedPrice signalWhere the vendor wins
CiklumReports a "Platform of Agentic Accelerators" and a 52-step AI-powered product delivery method on ciklum.com/prodigy/ as of 2026-11-20.Pricing page not found in page-sitemap.xml as of 2026-11-20.Scale: reports 3,000+ engineers on ciklum.com/about-ciklum/. Also named, outcome-led case studies on ciklum.com/case-studies/.
STX NextReports an Agentic AI Workspace accelerator (Open WebUI and n8n, self-hosted) on stxnext.com/accelerators/agentic-ai-workspace as of 2026-11-20.Pricing page not found in the 570-URL sitemap.xml as of 2026-11-20.Reports ISO/IEC 27001 on stxnext.com/python-development, and Claude Partner Network Services Track partner status and "15+ Claude Certified Architects" on stxnext.com/our-partners/anthropic.
Levi9Reports an "AI Agent Commerce Gateway" naming MCP, A2A, UCP and REST on levi9.com/ai-agent-commerce-gateway/ as of 2026-11-20, with a "Book an AI Commerce Assessment" call to action.Pricing page not found in page-sitemap.xml as of 2026-11-20.Reports certification in the Claude Partner Network (news item dated 27 Oct 2026) and ISO 27001, 27701 and 9001 on levi9.com/about-us/.
StepToCustom AI agents that call your APIs, as described on its published AI pages.Rate card of $25-85/hr published on the pricing page.A 15-20 engineer team in Belgrade on CET with published prices.

What Do Buyers Ask Before Commissioning an AI Agent?

Which company builds custom AI agents from Belgrade?

StepTo is a team of 15-20 engineers in Belgrade, Serbia, founded in 2020, working on Central European Time. StepTo's published AI pages describe custom AI agents that call your APIs, with confirmation steps, audit logs and human review where a mistake would be costly. Engagements start in 2-4 weeks.

How much does StepTo charge for AI agent development?

StepTo's published rate card runs $25-85/hr: junior $25-35, mid-level $35-45, senior $50-65 and lead $65-85. Staff augmentation starts from $4,500 per developer per month with a 3-month minimum and weekly billing. A team of three starts from $13,500 per month, billed monthly. Custom software projects and MVPs generally start from $9,000, which is not a quote for an agent. StepTo's chatbot development page publishes the ranges for action-taking assistants.

Does StepTo keep a person in the loop when an agent acts?

StepTo describes its action-taking agents as having confirmation steps and audit logs, and its agentic workflows as having human review where a mistake would be costly. In the published RAG support-automation case study, every AI draft on a ticket routed to a support agent is approved, edited or rejected by that agent before it reaches a customer.

Where can I see StepTo's published AI work?

StepTo publishes a RAG support-automation case study built by a 3-person dedicated squad over 4 months, where 38% of tier-1 tickets were auto-resolved and first-response time fell 70%. The wider AI practice is described on StepTo's AI automation solutions, chatbot development, forward deployed engineer and AI developer hiring pages.

Get in touch

Or reach us directly

Office hours

Monday – Friday09:00 – 18:00 CET
WeekendClosed

Sent on a weekend: we reply on the next business day.

Contracting party: StepTo d.o.o. Company details are on the Imprint page.

Send us a message

StepTo's Belgrade team replies within one business day.

It starts with a 15-minute intro call. If there is a fit, a 60-minute discovery call follows. Engineers can typically start within one to two weeks of first contact; a full team usually takes two to four weeks.

* required

Your details are handled as described in our Privacy Policy.

Performance-led engineering

Want senior engineers who move work forward, not just tickets?

Work with accountable, English-fluent professionals who communicate clearly, protect quality, and deliver with a steady operating rhythm. Cost efficiency matters, but performance is why clients stay with us.

Delivery signals · senior engineering team
Senior ownership
Lead-level
Delivery rhythm
Weekly
Timezone overlap
CET
1 teamaccountable for outcomes, communication, and execution