What coding agents can do
A coding agent is different from suggestions in an editor. A suggestion completes a line, while an agent takes the whole task: it reads the project, changes several files, runs the tests and repeats until the tests pass.
These agents share one loop. Anthropic describes it like this: gather context, take action, verify the result, repeat. On 29 September 2025 the company renamed the Claude Code SDK to the Claude Agent SDK, and outside services connect to it through the MCP protocol.
On 11 March 2025 OpenAI released the open Agents SDK with handoffs between agents, guardrails and tracing, that is, a record of every step the agent takes. For code, tracing is especially useful: it shows which commands the agent ran and why.
Most often an agent gets standard work: a draft function, tests, moving code to a new version of a library, analyzing someone else’s error from a log. The more a decision depends on agreements inside the company that are not in the code, the more changes a person makes afterward.
Hence the word vibe coding: development where a person describes the task in words and AI writes almost all the code. For a prototype this works; for a system with money and client data, not without review.
How the cost of development changes
An agent cuts the time to write code, but not the time to decide what to write or to check it. In a project budget this shows up as follows: standard parts get cheaper, and the share of work by the architect and the reviewer grows.
Trust in the result is still lower than use. In the same Stack Overflow survey, 46% of developers do not trust the accuracy of AI tools, and 33% do.
There is a downside too: code appears faster than people can read it. If the team accepts the agent’s changes without review, the savings on writing go into fixing errors after release.
Count the benefit by tasks, not by lines of code: how many tasks the team closed per month before agents and how many after, with the same number of errors after release.
For a client the conclusion is simple. A standard agent or integration is built from open blocks, and coding agents speed up that building. If you are charged for such work as for unique development, ask what the price is made of, and pay for a result against a metric, not for hours.
What an agent costs on a builder platform and as a turnkey project is covered on our page on the cost of an AI agent.
Code review
The agent’s code looks finished, and that is the risk. In the Stack Overflow 2025 survey, 66% of developers named AI solutions that are “almost right, but not quite” as their main problem.
Security is a separate topic. In July 2025 Veracode tested more than 100 models: 45% of AI-written code samples failed security tests, and for Java the failure rate reached 72%. New and large models wrote more working code, but their security did not improve.
1. People write or approve the tests before the agent starts on the code: a test fixes what counts as correct behavior.
2. Every change by the agent goes through review, that is, another developer reads the code, as with a change by an employee.
3. Before merging into the main code, changes are checked with a vulnerability scanner.
4. The agent is forbidden to change tests to make them pass. Reviewers look for such a change first.
The agent delivers changes in small parts so that each can be read in one sitting. A huge change of hundreds of lines is reviewed worse, and errors in it slip through.
The acceptance metric is the same as for people: how many of the agent’s changes were accepted without edits and how many errors were found after release.
Security and access
A coding agent runs commands on a developer’s computer or on a server. An error in a command or a harmful instruction in a file it reads can delete data or send keys outside.
So set up rights before the first task. For example, according to the Claude Code documentation, in manual mode the agent starts with read-only rights and asks for permission before it changes files or runs tests or commands. The user is responsible for checking the proposed code and commands.
Rules for a company: the agent works in a separate folder or container with no access to the live client database, it gets no keys to production systems, and every action is logged.
A person approves new libraries the agent proposes: each library is someone else’s code inside your program. Outside services connect through MCP servers, and the Claude Code documentation advises installing only servers from providers you trust, or writing your own.
Source code and client data are not sent to an outside model provider without a company decision. If they must not leave the company, the agent runs on an open model inside the company; read more on the page about a local agent.