Skip to Content
AI AgentsOverview

AI Agents

An agentic coding tool does more than answer questions. Give it a task in your repository and it reads files, writes patches, runs commands, reads the output, and iterates until the task is done — asking your permission along the way. The model doing the thinking is supplied by an API, and Devotel AI is a drop-in choice: every tool below talks to https://api.devotel.com/v1 with kimi-k3.

Verified tools

Each of these has been run against api.devotel.com with kimi-k3, end to end, on a real task.

ToolRuns whereStatus
OpenHarnessTerminal✅ Verified — recommended
OpenCodeTerminal✅ Verified
Cline & Roo CodeVS Code (and forks)✅ Verified
AiderTerminal✅ Verified

Tools we have looked at and cannot yet recommend — Cursor’s custom-API agent mode, OpenAI’s Codex CLI, Claude Code, Copilot BYOK — are covered on Compatibility notes, with the reason for each.

How an agent uses the API

An agent turns one instruction into a sequence of /v1/chat/completions calls. Each step sends the whole working context — your instruction, the files it has read, the commands it ran and their output, its own previous reasoning — and gets back either a tool call to execute or an answer. The context grows as the task progresses.

As a rough guide, from the runs behind this section:

Typical for one medium task
Requests~10 sequential /v1/chat/completions calls
Tokens40,000–80,000 total
ShapeInput grows each step; output is dominated by reasoning

Two consequences worth planning for:

  • kimi-k3 is a reasoning model. It spends output tokens thinking before it answers, and those reasoning tokens are billed as output. An agent that thinks its way through ten steps spends far more on output than the size of its final patch suggests. See Models.
  • Cost tracks the number of steps, not the length of your instruction. “Fix the failing test” is three words and can easily be fifteen requests.

Watch the real numbers on your usage dashboard  — agent traffic shows up there like any other API traffic.

Reasoning effort is your main cost lever

An agent loop makes many small calls, and most of them do not need deep thought — deciding which tool to call next is not the same problem as designing a schema. Because reasoning tokens are billed as output, at 5× the input rate, effort is usually the largest number you control:

  • reasoning_effort: "none" or "low" for tool-selection turns, routing and short extractions.
  • "high" or "max" for the planning turn that actually needs it.

Set a conservative default on the key (see Authentication) and raise it per request for the turns that earn it. Measure with usage.completion_tokens rather than guessing — it always includes reasoning, whereas the completion_tokens_details.reasoning_tokens breakdown is not populated on every response. On a single arithmetic question the gap between none and high was 8 tokens against 53; over a ten-step loop that compounds.

Rate limits

A new key carries no effective rate limit — the defaults sit at the ceiling of what the limit columns accept. That is a default, not a guarantee: an administrator can assign a specific requests per minute, tokens per minute, or monthly token budget to any individual key. Agent traffic is the most likely to meet one, if you:

  • run several agents, or several repositories, on one key
  • let an agent work on a very large context, where one step alone can be tens of thousands of tokens
  • share a key across a team

The X-RateLimit-* headers on every response report the limits your key actually carries.

If you hit 429, the response carries Retry-After and the X-RateLimit-* headers so a well-built harness backs off on its own — see Rate Limits & Errors. If you hit it routinely, ask your Devotel administrator to raise rpm_limit and tpm_limit on your key; both are per-key settings they can change without issuing you a new key.

Give each agent its own key. Separate keys keep the usage dashboard readable, let an administrator raise limits for the agent without touching your other integrations, and mean revoking one thing does not break the rest.

Before you start

  1. Create a key at chat.devotel.com/keys .
  2. Confirm it works before pointing a tool at it:
export DEVOTEL_API_KEY=dvt-YOUR_KEY curl https://api.devotel.com/v1/chat/completions \ -H "Authorization: Bearer $DEVOTEL_API_KEY" \ -H "Content-Type: application/json" \ -d '{"model":"kimi-k3","messages":[{"role":"user","content":"reply with OK"}]}'
  1. Pick a tool and follow its page. Then start the agent in a repository with a clean git status, so its edits are easy to review — and revert.