Loading the page.
Documentation
Journal
Frontend
Backend
Admin
CI/CD
Loading the page.
Card · concepts · updated 17.08.2026
concept · category for every tool in the table
AI agent - a program built on a neural network that gets a whole task, not a question. The agent plans its own steps, calls tools (edits files, runs commands, browses the web), checks the result, and brings back a finished result. A chat assistant answers and waits for the next question; an agent works until the task is closed. All 51 tools in our table are agents.
51 agents
In the registry
30 of 51
Call themselves agent
60–70% of tasks
Decide on their own
Paths exist
From Russia
task (“fix a bug”)
│
▼
┌-- plan the next step
│ │
│ ▼
│ tools: files · terminal · browser
│ │
└──── check ──── done ── result
# the model runs tools in a loop until the task is brought to a result▌What it gives you
What it is not
What we call it
We call it: AI agent. Also known as: AI agent, agent, coding agent, coding agent. Didn't stick: AI assistant as a category, autonomous assistant.
Verdict vibecoding.tech · 2026-08-17 · full glossary
Simon Willison's formula, which the industry has converged on: an agent - an LLM running tools in a loop toward a goal. Three details and a check:
Families · Forms - where the agent lives. One mechanism inside, six habitats; the dark tag in the table - the tool's native form.
| Type | What it's for | Today |
|---|---|---|
| Terminal | lives in the console, digs through the whole project | Codex, Claude Code, Gemini CLI - the engineers' form |
| Editor | agent inside the IDE, code in plain sight | Cursor, Kiro, Junie - the most crowded form |
| Desktop | its own window: sessions and tasks without a terminal | Codex and Claude Code as apps |
| Cloud | background executor: give it a task and leave | Devin, Jules, ChatGPT Agent |
| Builder | described it - got a website; the code is hidden | Lovable, Bolt, v0: avoid the word «agent» |
| Browser | browses the web and completes the whole task | Comet, Dia, Opera Neon |
The word «agent» predates the current wave and has already burned out once. The timeline explains why it won on the second try.
2021
GitHub Copilot
the first mass-market AI for code calls itself an «AI pair programmer»: suggestions alongside a person, not tasks
2023
AutoGPT
the first agent wave: 100 thousand stars in months and almost zero shipped products - the agent checked itself and looped
2024
Devin and the «AI engineer» frame
Cognition calls Devin «the first AI software engineer» - the market is trying to sell an employee replacement, not a tool
2025
Terminal wave
Claude Code and Codex make «coding agent» a working category; Willison records the definition: an LLM with tools in a loop toward a goal
2026
Umbrella term
our count as of 17.08: 30 of the table's 51 tools use agent/agentic in their self-description; the dissenters split among «harnesses», «platforms», and «assistants»
Choosing an agent means choosing a form and a benchmark. Honestly: start not with «the most powerful», but with the form where you can comfortably verify its work.
Self-description census
30 of 51 tools in our registry include “agent” or “agentic” in their self-description
Measured 08/17/2026 using vendors' official sites, docs, and READMEs. Other terms: “platform” and “builder” for builders, “harness” for terminal newcomers, “assistant” for RU vendors.
vibecoding.tech registry · 2026-08-17
Agents and code (2026)
Flagship setups solve 60–70% of real engineering tasks autonomously
Our anchor benchmark, DeepSWE v1.1, snapshot 07/13/2026: Codex + GPT-5.6 Sol - 72,7%, Claude Code + Fable 5 - 70%. So “which agent is better” is a measurable question, not a matter of taste.
vibecoding.tech index · 2026-07-13
From Russia
Paths exist: open agents (OpenCode, Cline, Qwen Code) can be installed freely, local assistants work directly
Western flagships (Claude Code, Codex) require overseas payment for a subscription or API - through intermediaries and ruble aggregators (markup ~1.5–2×).
Russia access (our guide) · 2026-08-17
There is maximum hype around agents in both directions. A breakdown of the most persistent claims:
«An agent is a chatbot with a new name»
the difference is measurable: a chat produces text, an agent changes files and runs commands. On our anchor benchmark, agents complete 60–70% of real engineering tasks without a person
«This is AutoGPT, we've been through this already»
the 2023 wave checked itself and looped; 2025–2026 agents rely on external verification - tests, CI, review - and ship real PRs
«An agent and a model are the same thing»
a model is the brain, an agent is a brain with hands and a loop. One Fable 5 lives in Claude Code, Cursor, and cloud agents; you need to compare the «agent + model» combination
«Agents replace programmers»
they replace part of the work: routine goes to the machine, while task setting, verification, and architecture become more valuable. Merge authority stays with the person
«You cannot use agents from Russia»
open agents install freely, local assistants work directly; Western flagships - with foreign payment or through intermediaries
Every vendor calls its thing differently. Our census of the self-descriptions of all 51 tools in the table (17.08.2026) comes down to six frames:
| Term | What it is | Example |
|---|---|---|
| ·Coding agent | dominant frame: an agent for coding tasks | Cursor, Codex, Cline, Junie, Jules, Droid |
| ⊃Agentic IDE | an editor or environment where the agent is built into the workflow | Kiro, Windsurf, Warp, Firebase Studio |
| ⊃⊃Harness | the setup around the model: tools, loop, interface | DeepSeek Harness, Grok Build, Pi |
| ⊃⊃⊃AI software engineer | the «employee replacement» frame, not a tool | Devin; Trae promises a «10x engineer» |
| ⊃⊃⊃⊃AI pair programmer | the 2021 frame: suggestions alongside a person | GitHub Copilot, Aider |
| ⊃⊃⊃⊃⊃Assistant / helper | the choice of all Russian big tech: agency is a mode, not the essence | GigaCode, SourceCraft, Kodify |
A neural network-powered program that receives a whole task, not a question. An agent plans steps on its own, edits files, runs commands, checks the result, and delivers the finished work. A chat answers and waits for the next question; an agent works until the task is closed.
By doing and iterating. ChatGPT produces text, then you take over. An agent acts: it changes project files, runs tests, opens websites, and repeats the “step - check” loop until it gets the result. It can use the same model as a chat.
The scaffolding around a model: tools, a loop, and an interface. The model - the brain, the harness - the rest of the agent's body. The word moved from engineering jargon into self-descriptions in 2026: DeepSeek Harness explicitly calls itself an “agent harness,” as do Grok Build and Pi.
The word beat its rivals on coverage: “IDE” does not cover the terminal, “assistant” sounds like the previous era of suggestions, “harness” is jargon. Our measurement from 08/17/2026: 30 of 51 registry tools include “agent” or “agentic” in their self-description, and the share is growing - CodeRabbit and Augment changed their positioning to “agentic” in 2026.
Our verdict (08/17/2026): AI agent. A count of the self-descriptions of 51 registry tools: 30 include “agent” in their self-description, and it is the only word that spans all six formats - from terminal to browser. “Assistant” is the choice of RU vendors and the previous paradigm of suggestions; “harness” is the engineering word for the scaffolding around a model, part of an agent rather than a synonym.
For code, it is measurable: our index is led by Codex (72,7% on DeepSWE) and Claude Code (70%, 78,7% on RuBench). For websites without coding - builders like Lovable and Bolt, which are not included in code benchmarks. There is no universally “best” option, only a ranking for the task.
Open agents (OpenCode, Cline, Aider) are free; you only pay for the model: a typical task on a flagship costs a few dollars. Vendor subscriptions - $20–200 per month depending on limits. Average task price in our measurement: from $0,55 (Cursor) to $11,75 (Claude Code with Fable 5 at maximum).
Yes, but differently. Open agents can be installed freely; the question is model access: DeepSeek and local options (GigaCode, SourceCraft) work directly, while Western flagships require overseas payment or intermediaries with a markup of ~1.5–2×.
Related concepts
Loopsthe next level: an agent wrapped in a trigger, verification, and a stop condition
MCPthe socket through which the agent connects tools
Neural networkthe foundation: an agent is a neural network given hands
More on the topic
Entry · ai-agent · facts · 7 (source and date for each) · card edited 2026-08-17 · facts verified 2026-08-17
Where to go next
New cards, price and limit changes, ranking shifts - by email, no spam.