A coding agent is an AI agent whose tools are your editor, file system and terminal.
How it works
The agent gets a task in plain language and works in a loop. Each turn, the model either answers or returns tool calls such as “read this file”, “apply this edit” or “run the tests”. The tool host runs them, appends the output and calls the model again.
Claude Code’s docs describe it as “an agentic coding tool that reads your codebase, edits files, runs commands, and integrates with your development tools.” Aider, Cline and similar tools follow the same loop with different interfaces and permission models.
The SWE-agent paper (Yang et al., 2024) argues that agents are a new kind of user that needs its own interface, an “agent-computer interface”, in the way people need an IDE. With their interface, the authors report a pass@1 rate of 12.5% on the SWE-bench benchmark, which they describe as state of the art at the time.
A typical run, as requests to the model:
| Step | Model output | Tokens sent as input |
|---|---|---|
| 1 | run_tests() |
system prompt + tools + task |
| 2 | read_file("src/parse.py") |
step 1 + test output |
| 3 | edit_file(...) |
step 2 + file contents |
| 4 | run_tests() |
step 3 + diff |
| 5 | “Fixed; tests pass.” | step 4 + test output |
Input grows every step, because the whole history is resent.
Why it matters
- Cost. Long tasks mean many requests with large, mostly repeated inputs. The run cost post works through real numbers, and cutting coding agent costs lists the levers.
- Context. Long sessions fill the context window. Agents use context compaction and subagents to keep it in check.
- Model fit. The model must follow tool schemas reliably over many turns. See the best open model for coding agents.
- Safety. An agent that runs commands can do damage. Keep approval on for destructive actions and treat repository content as untrusted input (prompt injection).
On BroadNet Inference
Coding agents that accept an OpenAI-compatible base URL can call BroadNet at https://inference.broadnet.ai/v1. Setup guides exist for Aider, Cline and others. BroadNet’s docs list tool calling for GLM 5.3 and Qwen3.8 27B.
On pay-as-you-go, a key allows 30 requests per minute and 4 in flight. The flat-rate plans (Full Moon, Orbit, Meteor) cover DeepSeek V4.1 Flash with no token meter. A session runs one request at a time, which fits one agent. If you run agents in parallel, or an agent that starts subagents, buy one session for each request that runs at the same time. The sessions guide shows the math.