Agentic coding: what it is and how it differs from autocomplete

Agentic coding is a way of working where the AI receives a whole task and brings it to a result itself: it reads the files it needs, makes edits, runs tests and fixes its own mistakes from their output. The human sets the goal and checks the outcome. Below we cover how this differs from editor suggestions, what an agent consists of, which tools exist and what an hour of its work costs via API.

Three generations of coding help

It is useful to tell three levels apart.

1. Autocomplete. The editor suggests the continuation of a line or block. The model sees the neighbouring code, the decision is yours, and each accepted suggestion is one human action.
2. Chat. You describe a task, the model replies with a code fragment, you copy it into the project and run it yourself. The context is mostly what you pasted into the window.
3. Agent. The model gets access to tools: reading and searching files, editing, a terminal. It decides what to open and looks at the run result itself.

The difference is not the "smartness" of the model but a closed feedback loop. In the first two cases feedback goes through you; in the third the model gets it directly: the test failed, so the edit is wrong.

Agentic coding: what the loop looks like inside

Every agent is built around the same loop:

1. The model receives a task and a list of tools (read a file, search the project, edit a file, run a command).
2. It replies not with text but with a request to call a tool.
3. The wrapper program (called a harness) executes the call and returns the result into the conversation.
4. Steps 2-3 repeat until the model decides the task is done or you stop it.

Two practical conclusions follow. First, the quality of the result depends not only on the model but on the harness: how it searches the project, what it allows to run, how it compresses history. More in what an agent harness is↗. Second, on each step the model is sent the whole accumulated conversation, and that determines cost.

Which tools exist

All of them follow the loop described above and differ in where they live.

• Terminal agents: Claude Code, opencode, Codex CLI, Aider and others. They run next to the editor and are handy on a server over ssh. Connecting Claude Code to the gateway is described in the docs↗.
• Editor extensions: Cline, Kilo Code and similar (Roo Code, according to its developers, was shut down on May 15, 2026). Edits are visible right in the files. Overview: VS Code extensions for working with AI↗.
• Editors with a built-in agent: for example, Cursor.

Most of them accept any compatible API, so you can choose the model separately from the tool. It is a matter of configuration: an address and a key.

What an agent does well and where you need control

Tasks where feedback can be formalized as a command suit it best: fix a failing test, add an endpoint modelled on existing ones, update a dependency and repair the build, write tests for a function, port a module. All of these have a check the agent can run itself.

Control is needed where there is no check or it is incomplete: architectural decisions, security, data migrations, anything that touches production. Practical rules:

• work in a separate git branch and read the diff before committing;
• let the agent run tests and read files, and confirm dangerous commands by hand;
• do not give the agent access to production secrets and servers;
• phrase the task together with a way to verify it.

A related approach, where the human trusts the agent more and barely reads the code, is called vibe coding; for production projects it is riskier.

What an hour of agent work costs via API

Payment is per token, so the calculation is simple. Assumptions (illustrative, for the calculation): the agent makes one request per minute, so 60 requests an hour; on average each request uses 50,000 input tokens (accumulated conversation and files) and 1,000 output tokens. Total per hour: 3M input tokens and 0.06M output.

Formula without cache: input (M) × input price + output (M) × output price. Zukko prices as of October 18, 2026, $ per 1M tokens:

This is an estimate without cache. Agent sessions benefit a lot from prompt caching: in the Anthropic format (Claude Code, Anthropic SDK) the gateway sees cache reads and writes and bills them separately. For claude-sonnet-5-5 assume 90% of input is read from cache and the other 10% is written to cache: 2.7 × 0.05 + 0.3 × 0.5 + 0.06 × 2 = 0.135 + 0.15 + 0.12 = $0.405 per hour. For GPT models only the fact of a cache read is visible, for others it depends on the provider; details in calculating agent cost with cache↗.

The real figure will differ, because it depends on project size and the agent's habits. A cheap model may need more steps, and an expensive one may finish in fewer; the strength of the models by the Artificial Analysis index in our catalog: Opus 5.5 is 58, Sonnet 5.5 is 48, GPT-6.1 Sol is 52, DeepSeek V4.1 Flash is 39. Test on your own tasks rather than by the table. Current prices: GET https://api.zukko.pro/v1/models.

Where to start

Pick one tool and one small task with a check. Connect a key from the dashboard↗, set a daily limit in dollars on the key and look at spend in the "Usage" section after the first session. Access from Russia goes through the gateway; payment by SBP, MasterCard card or cryptocurrency.

FAQ

Does agentic coding replace the developer?

No. The agent performs tasks with a clear check, but setting the goal, architectural decisions and responsibility for the result stay with a person. A good analogy is a fast intern whose work you read in review.

How is an agent different from an ordinary chat with a model?

Access to tools and a closed loop. A chat answers with text, while an agent opens files, edits them and runs commands itself, seeing the result.

Why is an agent more expensive than a chat?

Because it makes many requests per task and each carries the accumulated context. Cache, short sessions and precise instructions about which files to look at help.

Do I need a special model?

You need a model that calls tools reliably. We do not claim in advance which one does better; compare two or three models from the catalog on the same task.