AchiralAchiral

Docs · LLMs and developers

Reviewed2026-07-22Version3.12.1

Agents quick start

Create an agent and call it through the inference runtime.

This guide walks through one simple goal: create an agent for a recurring workflow, then call it through the runtime.

1. Pick one recurring workflow

Start with a workflow that repeats often enough for memory to matter.

Good first agents usually have:

  • a clear name;
  • a clear responsibility;
  • a narrow boundary; and
  • an owner who can judge whether the memory stays useful.

Example:

release-readiness-agent

2. Create the agent in the app

Use the Agents area in the organization UI to create the agent.

At creation time, choose:

  • whether it is durable or ephemeral;
  • the responsibility name;
  • the client or account boundary if your workflow needs one; and
  • any initial project or scope values you want to carry into runtime calls.

3. Use a Context Access Token

Use a Context Access Token that can call the inference runtime.

export ACHIRAL_TOKEN="acm_..."
export ACHIRAL_AGENT="release-readiness-agent"

4. Call the runtime

You can call the agent through the same OpenAI-compatible runtime shape used elsewhere in the product.

curl https://your-org-slug.achiral.ai/v1/chat/completions \
  -H "Authorization: Bearer $ACHIRAL_TOKEN" \
  -H "Content-Type: application/json" \
  -H "X-Achiral-Agent: $ACHIRAL_AGENT" \
  -d '{
    "model": "chiro",
    "messages": [
      {
        "role": "user",
        "content": "Summarize what I need for the next release-readiness review."
      }
    ],
    "temperature": 0.2,
    "max_tokens": 800
  }'

You can also use the route-based agent endpoint documented in API reference.

5. Keep the scope narrow

Before you rely on the agent in production, confirm that:

  • the right agent identity is present on every request;
  • project or workflow scope is added when needed; and
  • the agent is not being used across unrelated customers or responsibilities.

6. Review the result in practice

Start with one workflow, run it a few times, and check whether the recalled context is actually helping. If it is not, narrow the boundary or adjust how the workflow is split into agents.

Continue from here