Env API

The environment orchestrator (/v1/env/*, /tasks, /episodes/*) runs on the private network and is not published at agi.akhara.ai.

Audience	How to call
Most clients	Use the Evaluations API on `https://agi.akhara.ai`
Operators	Call the orchestrator on the private host (internal URL / VPN / SSH tunnel)

Paths below are relative to that private orchestrator base. Schemas are per-environment; examples use the mobile shopping environment.

Capabilities

GET /v1/env/capabilities

{
  "env_id": "amazon_clone_mobile",
  "env_version": "1.0.0",
  "schemas": {
    "observation": "amazon.mobile.obs.v1",
    "action": "amazon.mobile.action.v1",
    "step_result": "amazon.mobile.step_result.v1",
    "verifier_result": "amazon.mobile.verifier_result.v1"
  },
  "actions_supported": ["tap", "type_text", "scroll", "select", "copy", "paste", "back", "finish"],
  "limits": {
    "max_steps_per_episode": 80,
    "max_action_retries": 2,
    "max_text_len": 200
  }
}

Task bank

GET /tasks
GET /tasks/{task_id}

GET /tasks/{task_id} includes a harbor_reset object (legacy field name): the exact payload to pass into reset. Public equivalents:

curl -s https://agi.akhara.ai/evaluations/tasks
curl -s https://agi.akhara.ai/dashboard/tasks/vu.checkout.t098

Reset

POST /v1/env/reset

{
  "episode_id": "ep_demo_001",
  "seed": 18421,
  "fixture_id": "fixture1",
  "max_steps": 38,
  "task_spec": {
    "task_id": "vu.checkout.t098",
    "goal": "Find me a red hoodie and checkout.",
    "success_criteria_id": "SC_CHECKOUT",
    "verifier_id": "VU-CHECKOUT",
    "target_query": "hoodie",
    "target_sku": "B07HOODIE02",
    "target_variant_name": "Red",
    "matching_policy": "deterministic_sparse_reasoned",
    "constraints": {
      "clean_cart_required": true,
      "must_be_in_stock": true
    },
    "require_needs_confirmation": true,
    "profile_id": "U_001"
  }
}

Returns episode_id + initial observation.

Step

POST /v1/env/step

{
  "episode_id": "ep_demo_001",
  "step_id": 1,
  "type": "click",
  "params": { "point_2d": [540, 960] }
}

Response highlights:

{
  "schema": "amazon.mobile.step_result.v1",
  "episode_id": "ep_demo_001",
  "step_id": 1,
  "accepted": true,
  "observation": { "schema": "amazon.mobile.obs.v1", "step_id": 1 },
  "verifiers": [
    {
      "verifier_id": "VU-CHECKOUT",
      "status": "IN_PROGRESS",
      "reward": { "task": 0.0, "safety": 0.0, "efficiency": -0.01, "format": 0.0 }
    }
  ],
  "done": { "is_done": false, "reason": null, "final_score": null },
  "reward": 0.0,
  "truncated": false
}

Convenience episode API

Used internally by the eval wrapper; also available on the private orchestrator:

Method	Path
`POST`	`/episodes/start`
`POST`	`/episodes/{episode_id}/step`
`GET`	`/episodes/{episode_id}/status`
`DELETE`	`/episodes/{episode_id}`

POST /episodes/start body:

{
  "task_id": "vu.checkout.t098",
  "max_steps": 38,
  "seed": 18421,
  "episode_id": "ep_abc123def0"
}

Returns a flatter observation including screenshot_b64 and ui_elements when available.

Export episode

GET /v1/export/run/{episode_id}

Returns application/zip after the worker has synced (POST /v1/sync/run). Contents typically include run.json, steps.jsonl, per-step JSON, and optional sqlite.db / manifest.json.

Workers (ops)

Method	Path
`POST`	`/v1/workers/register`
`GET`	`/v1/workers/list`
`GET`	`/v1/workers/health`
`POST`	`/v1/workers/{worker_id}/set_health`

Clients running evaluations do not need these unless they operate the fleet.

Getting started

Mechanics

API reference

Scores and trajectories

Capabilities

Task bank

Reset

Step

Convenience episode API

Export episode

Workers (ops)

​Capabilities

​Task bank

​Reset

​Step

​Convenience episode API

​Export episode

​Workers (ops)

Capabilities

Task bank

Reset

Step

Convenience episode API

Export episode

Workers (ops)