Skip to content

Can you trust an AI agent inside your product?

Mykhailo PoliarushUpdated 5 min read

In 2026, AI agents are moving from demos into products, and reliability, not raw capability, decides which ones survive contact with real users. In August 2025, Gartner forecast that 40% of enterprise applications will include task-specific AI agents by the end of 2026, up from less than 5% in 2025, yet only 16.0% of developers in the Stack Overflow Developer Survey 2026 say their AI tools handle complex tasks very well. This article covers agents inside your product: where adoption stands, what goes wrong, what they cost, which protocols matter and how much supervision they need.

How far have AI agents moved into products?

Further than most tools, not as far as the forecasts. In the Stack Overflow Developer Survey 2026, 26.2% of 17,464 respondents use AI agents or automated workflows at work, against 65.9% who use AI coding assistants or coding agents. Investment is similarly cautious: in a January 2025 Gartner poll of 3,412 webinar attendees, 19% said their organization had made significant investments in agentic AI, 42% conservative ones and 31% were waiting or unsure.

Where agents do ship, they increasingly face customers. In LangChain's State of Agent Engineering survey of 1,340 practitioners, customer service was the most common primary use case at 26.5%, followed by research and data analysis at 24.4%. LangChain sells agent tooling, so its audience leans toward teams already building agents. For production rates, observability and eval adoption across the industry, see our report on the state of AI engineering in 2026.

What are the main challenges of putting an agent in a product?

1. Inconsistency. An agent that solves a task once may fail the same task on the next run. The tau-bench study (June 2024) measured this with a metric called pass^k, the chance of succeeding on all of k attempts: the best function-calling agent at the time succeeded on under 50% of tasks, and its pass^8 in a retail scenario was below 25%. Those models are older now, but the lesson stands: a product agent should be judged on repeated runs, not a lucky demo.

Developers feel the gap. In the Stack Overflow Developer Survey 2025, 87% of respondents were concerned about the accuracy of AI agents and 81% about the security and privacy of data. In 2026, 38.8% rated their AI tools "good, but not great" at complex tasks and 24.0% rated them bad or very poor, according to Stack Overflow.

Developers who strongly agree with each statement about AI agents

Concerned about accuracy
57.1%
Concerned about data security and privacy
56.1%
Cost of agent platforms is a barrier
25.4%
Integration with existing tools is difficult
16.5%
Learning to use agents takes significant effort
15.5%
IT or InfoSec rules block agent tools
13.8%

Source: Stack Overflow Developer Survey 2025, AI agents (28,930 responses, July 2025)

2. Cost that grows with autonomy. Agents loop: they plan, call tools, read results and try again. In Anthropic's account of its multi-agent research system, agents used about 4 times more tokens than chat interactions and multi-agent systems about 15 times more. Anthropic found that token usage alone explained 80% of the performance variance on a browsing benchmark, so spending more often buys quality. These are a model vendor's own figures from its own system, but the direction is a useful planning rule: multi-agent designs pay off only when the task is worth it.

Approximate token use compared with a chat interaction

Chat interaction
1x
Single agent
4x
Multi-agent system
15x

Source: Anthropic (model vendor), multi-agent research system, own data, 2025-06-13

Speed is the other cost. LangChain's survey ranks latency as the second biggest barrier to production at 20%, behind quality, and notes that raw spend is cited less often than in previous years.

3. A wider attack surface. An agent that reads user input and calls your APIs can be steered by that input. The OWASP Top 10 for LLM Applications 2025 lists prompt injection as risk number one and "excessive agency", meaning more permissions or autonomy than the task needs, as number six. Among enterprises with 2,000 or more employees, security is the second largest concern after quality, cited by 24.9% in the LangChain survey.

4. Projects that never pay back. Gartner predicts that over 40% of agentic AI projects will be canceled by the end of 2027 because of escalating costs, unclear business value or inadequate risk controls. The same Gartner release estimates that only about 130 of the thousands of vendors selling "agentic AI" offer real agents, a practice it calls agent washing.

Which protocols should an agent speak: MCP or A2A?

Usually both, for different jobs. The Model Context Protocol connects an agent to tools and data. Anthropic donated MCP to the Agentic AI Foundation, a Linux Foundation fund co-founded by Anthropic, Block and OpenAI, in December 2025, so it is no longer one vendor's standard. Agent2Agent (A2A) lets agents from different vendors discover each other and coordinate; the Linux Foundation took over A2A in June 2025, with support from more than 100 technology companies.

Builders already use them. Among developers who use agents, 38.3% use them to integrate with external agents and APIs and 34.4% for MCP servers, according to the Stack Overflow Developer Survey 2025. Among the 4,400 respondents who answered the orchestration question in the 2026 survey, LangChain leads at 30.2% and the OpenAI Agents SDK follows at 23.8%, while 18.9% orchestrate directly from the terminal. No framework dominates, which argues for keeping your agent logic portable.

How much human supervision does a product agent need?

Enough that a wrong action is caught before it costs money or trust. Teams that evaluate agents still lean on people: in the LangChain survey, 59.8% use human review and 53.3% use a model as a judge, and 62% have detailed tracing of each agent step and tool call.

In practice, supervision is a design decision per action:

  • Read-only actions (search, summarize, draft) can run freely and be sampled for review.
  • Reversible actions (create a ticket, update a draft) can run with logging and easy undo.
  • Costly or irreversible actions (refunds, emails to customers, deletions) need an explicit human approval.

What does a good product agent look like, and how do you start?

  • One defined job. Scope the agent to a task with a clear success criterion, not "an assistant for everything".
  • An eval set from real cases. Collect a few dozen real examples, score every prompt or model change against them, and track pass rates over repeated runs.
  • Least privilege. Give each agent only the tools and data its task needs, behind your existing auth.
  • A cost and latency budget. Set limits per run and alert when an agent loops.
  • Tracing from day one, so every failure can be replayed and turned into a test.
  • Start narrow. Pick one workflow where a person already checks the result, put the agent before that person, and measure how often they accept its output.

Want an agent in your product that you can trust?

SDCLabs builds AI agents into products: scoped to real tasks, connected to your tools and data, with guardrails, human approvals and an eval suite built from your own examples. Senior engineers run AI-assisted workflows and decide what ships, and automated tests from the founders of Codeception and CodeceptJS check every change. Engagements start in 1 to 2 weeks, and we reply within 2 hours on business days.

Tell us which workflow you want to automate. Related: AI agent benchmarking and AI testing.

Sources checked on 2026-11-17.

Frequently asked questions

How many developers use AI agents at work in 2026?

In the Stack Overflow Developer Survey 2026, 26.2% of respondents used AI agents or automated workflows at work, against 65.9% who used AI coding assistants or coding agents.

How much more do AI agents cost to run than a chatbot?

In Anthropic's multi-agent research system, agents used about 4 times more tokens than chat interactions and multi-agent systems about 15 times more. These are a model vendor's own figures.

What is the difference between MCP and A2A?

MCP connects an agent to tools and data and was donated to the Agentic AI Foundation in December 2025. A2A lets agents coordinate with each other and moved to the Linux Foundation in June 2025.

What are the main security risks of AI agents?

The OWASP Top 10 for LLM Applications 2025 ranks prompt injection first and lists excessive agency, giving an agent more permissions or autonomy than its task needs, as risk number six.

Related posts

Get an estimate for your product.