Agentic OS (Forward HIVE)

Build a bounded agent run with state, authorized tools, approval policies and usage controls.

Agentic OS is Forward's agent runtime offering, powered by Forward HIVE. It coordinates stateful runs, scoped tools, artifacts and configured human approvals. Use it directly through the API, or ask Forward to implement and operate a workflow on Managed Agents.

The configured deployment and acceptance tests determine which controls apply. A checkpoint does not itself guarantee that an external write is idempotent, and a matching citation does not prove that a conclusion is correct.

First agent request#

This template follows the Agent API delegation contract. It is not a live deployment or availability test. Onboarding supplies:

FORWARD_AGENT_BASE: the API origin, without a trailing slash.
FORWARD_AGENT_KEY: a bearer key authorized for your workspace.
An enabled agent_id, with its tools, access and approval policy agreed in advance.
Effective billing terms and request limits.

Use synthetic data first. In a private working directory, run umask 077 and create first-agent.json. Replace the agent placeholder before sending:

JSON

{
  "agent_id": "<enabled-agent-id>",
  "goal": "Review the synthetic supplier note and return a short summary with unresolved questions. Do not contact the supplier.",
  "context": {
    "supplier_note": "Synthetic example: delivery is proposed for 30 days after acceptance. The warranty period is unspecified."
  },
  "budget": {
    "max_acu": 2,
    "max_wall_clock_s": 60
  }
}

The budget values are illustrative limits, not a cost estimate or a performance promise. Confirm the enabled agent's permissions before sending; goal text does not replace a tool authorization policy.

Shell

curl --fail-with-body --silent --show-error \
  -X POST "${FORWARD_AGENT_BASE}/v1/agent/delegate" \
  -H "Authorization: Bearer ${FORWARD_AGENT_KEY}" \
  -H "Content-Type: application/json" \
  --data-binary @first-agent.json \
  --output run.json

A successful delegation returns HTTP 202 with run_id, thread_id, status and stream information. A returned delegation token is sensitive. Keep run.json private and read only the fields needed for inspection.

Inspect and approve#

Inspect GET /v1/agent/runs/{run_id} using the same bearer authorization. The run's event stream is available at GET /v1/agent/runs/{run_id}/stream; reconnect behavior and retention must be qualified for your deployment.

An interrupted run needs a decision under the agreed policy. POST /v1/agent/runs/{run_id}/resume accepts decision (approve or reject), an optional note, and nonnegative additional_acu. Review the proposed action before approving. Checkpoint recovery and external-system retries need their own acceptance cases.

Before expanding, review task completion, source support, tool permissions, approval behavior and recorded usage against the expected result.

Controls to qualify#

Control	Scope and acceptance evidence
State and recovery	Confirm persistence, interruption, expiry and safe retry behavior on the enabled runtime.
Data and tools	Validate workspace authorization, connector credentials, allowed actions and data retention. Identity headers alone do not establish authorization.
Human approval	Define the actions and budget thresholds that require a decision. Test both approval and rejection.
Source checks	Define supported sources, schema checks and unresolved-evidence behavior. Treat factual and professional review as separate acceptance criteria.
Model routing	Use account-enabled model IDs and qualify their result quality, latency and deployment requirements.

Commercial structure and ACUs#

Agent execution uses Agentic Compute Units (hive_acu). The pricing reference lists metered primitives and an illustrative worked example. It does not establish a universal USD conversion, starter allowance or pause-time exemption.

The account agreement defines monetary rates, effective date, included inference, tool charges, retained resources, retry treatment, overages and limits. When inference is included in an agreed ACU package, it must not also create a second token charge. Separate direct Grid calls use their own terms.

Compare API, implementation and managed responsibilities, or use the Grid model quickstart for direct model inference.