Docs
Capabilities

Agent runs

Run a server-side agent loop with tool calling and backtracking

POST /v1/agent/run runs an autonomous agent loop on the server. The agent works toward a task, calls the tools you allow, and can backtrack when a path stops paying off. You get the loop without building it client-side.

Parameters

  • task (required): the goal, 1 to 32000 characters.
  • tools: an allowlist of tools the agent may call.
  • max_steps: cap on loop iterations.
  • max_total_tokens: total token budget, at least 1024.
  • max_tokens_per_step: per-step token cap, at least 256.
  • reasoning_effort: none/minimal (fast), low (brief thinking), medium (thinking), high/xhigh/max (thinking, larger reasoning budget). Default medium.
  • stream: stream events over SSE. Default true.
  • confidence_threshold: minimum confidence the agent aims for.
  • compression: an object with enabled and trigger_tokens to compress context as it grows.
  • glassbox: glassbox options object.
  • refusal_mode: soft or strict.

Run with a JSON response

Set stream to false to get a single JSON body instead of an event stream.

curl https://api.ablatic.ai/v1/agent/run \
  -H "Authorization: Bearer $ABLATIC_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{
    "task": "Find the three largest cities in Austria and list their populations.",
    "max_steps": 8,
    "stream": false
  }'

The response carries an X-Run-Id header that identifies the run.

Stream events

With stream left at its default of true, the run emits Server-Sent Events. The event types are:

  • step: the agent took a step.
  • tool_result: a tool returned a result.
  • backtrack: the agent abandoned a path and went back.
  • content: output content.
  • done: the run finished.
  • error: the run failed.

Cancel a run

POST /v1/agent/run/{run_id}/cancel

Use the run_id from the X-Run-Id header of the run you want to stop.

Slots

Agent runs need an available agent slot. A request with 0 slots returns 403. Each running agent holds one slot, and if it spawns subagents, each running subagent holds one too, so a run that fans out can use several at once. A run that would exceed your slot limit returns 429.

Accounts get 2 agent slots by default.

Next