Agentic OS / Enterprise operating layer

From a business request to an accountable agent.

Describe the work, scope an approved pattern or bring an existing agent. Define the job, owner, tools, limits and tests with your team. Forward HIVE powers governed execution; authorized outcomes feed evaluated improvements and eligible Agent Learning.

Request Agentic OS API access Read runtime docs

Runtime contractAgent Protocol v1 + SSE

Thread state: thrd_* / thrd_autorun_*
Tenant boundary: X-Organization-Id · Vault ACLs
Verification: Citation & policy gates
Human control: Interrupt & resume envelope
Unit economics: hive_acu ledger metering

Execution engine

Forward HIVE

Stateful Agent Protocol threads with durable checkpointing and replay.

Tenant isolation

Scoped tool access

Access and isolation controls are confirmed for your deployment.

Safety & control

Verifier + HITL

Deterministic citation/policy gates and resumable approval interrupts.

Learning loop

Governed improvement

Authorized outcomes and calibrated feedback inform evaluated changes and eligible reward-based training.

[01 / CORE ARCHITECTURE]

Six capabilities to scope for your workflow.

Agree the behavior, connected systems and acceptance criteria for each deployment. Review the applicable controls and technical references below.

01 / ORCHESTRATION

Durable multi-step execution & checkpointed state

Persist workflow state and resume interrupted work under the recovery behavior qualified for your deployment. External effects require their own idempotency and reconciliation checks.

+Agree run budgets, timeouts and eligible model choices, then test recovery and uncertain external actions.

Versioned run state and checkpoints support recovery; the accepted profile defines persistence, retries and external-effect reconciliation.

02 / TOOL SANDBOX

Scoped tools, sandboxes & tenant controls

Scope which tools and data an agent can access. Identity, authorization, isolation and revocation are configured and verified for the agreed deployment.

+Grant read-only, staged-write, or approval-gated permissions per connector and environment.

Authenticated tool bindings and allowlists enforce the configured access boundary. Header propagation alone does not establish authorization.

03 / APPROVALS & HITL

Native human-in-the-loop interrupts & budget soft-pauses

Configure approval points and budget thresholds for the actions in your workflow. Qualify interruption, resumption and revocation before enabling those actions.

+Assign reviewers and define which actions require approval, which can run automatically and how exceptions are handled.

First-class `__interrupt__` state transitions (`status === "interrupted"`, `external_tool_budget_soft_pause`) that bind directly to the active `run_id` for scoped resumption; external side effects require their own recovery policy.

04 / PROVENANCE & SAFETY

Deterministic citation audits & OpenDetector guardrails

Check supported source references and output schemas, flag unresolved evidence and route exceptions for review. Citation checks do not establish the correctness of every claim.

+Select `warn_only`, `block_invalid_only`, or `block_all` gate policies per workspace tier and workflow criticality.

Terminal citation safety engine (`/reviews/citation_audit/{run_id}.json`) paired with OpenDetector AI and OpenDetector Injection checks at the retrieval boundary.

05 / MEMORY & VAULTS

Tenant-isolated vaults, document parsing & artifacts

Scope document ingestion, retrieval and artifact generation around your file types, access rules and retention requirements.

+Agree data rights, retention, export and deletion for files, deliverables and logs.

Dedicated `/v1/{org}/vaults/...` object storage and vector/lexical indexing with terminal attachment synthesis recovery.

06 / METERING & COST CONTROL

Unified Agent Compute Unit (ACU) metering & attribution

Track agent tasks, tool calls, browser sessions, sandbox seconds, and sub-agent delegations through one transparent unit of account—without double-charging included model tokens.

+Set hard ACU caps per department, workflow, or customer tenant with real-time burn telemetry.

Deterministic conversion table (`hive_acu`) reconciled against outbox usage dispatch and invoice-level token-inclusion guards.

[02 / SCOPED WORKFLOW WALKTHROUGH]

Inspect an illustrative document workflow.

Illustrative workflow: a supplier agreement moves through extraction, policy review and a human-approved handoff. The sample IDs, timings, ACU amounts and protocol snippets are examples, not a live customer run or an invoice.

STAGE 1 OF 06 · Customer System

01. Scoped Intake & Policy Binding

THREAD_INITIALIZED

A procurement compliance workflow is triggered with a 64-page supplier agreement, tenant headers, and an explicit 5.00 ACU run budget.

Illustrative output envelope

{
  "run_id": "run_01J9X8B7K",
  "status": "running",
  "policy": { "tenant_scope": "illustrative-policy", "external_write": "requires_approval" }
}

Example control and acceptance check

Illustrative policy check. Actual identity, permissions and data access must be qualified for the deployment.

Read the scoped Forlex reference

[03 / OPERATING MODELS]

Build with the API or ask Forward to implement and operate.

Choose who implements the workflow and who operates it. Your written scope confirms integrations, deployment, support and acceptance.

Your engineers build the agent graphs; Forward HIVE runs the state, sandboxes, approvals, and metering.

Agentic OS Runtime API

Product & platform teams shipping customer-facing or internal agents who want full code control without building orchestration plumbing.

Customer owns: Agent system prompts, tool definitions, and application UI; Connector credentials and domain business logic; Human-in-the-loop reviewer assignment inside your product
Forward owns: Durable thread execution, checkpointing, and interrupt/resume state; Isolated tool sandboxes, browser automation, and vault storage; Terminal citation verification, OpenDetector guardrails, and ACU telemetry
Commercial structure: Contracted ACU plan and usage. The agreement identifies included inference, external resources, support and any additional charges.

Explore details

Forward implements and operates the agreed workflow, integrations, evaluation and support scope.

Managed Agents

Enterprise and regulated teams deploying multi-step operational agents that require strict uptime, audit evidence, and continuous regression gates.

Customer owns: Domain policies, approval thresholds, and final sign-off on high-risk actions; Internal system-of-record permissions and data residency rules; Business KPIs and acceptance sign-off criteria
Forward owns: Managed HIVE runtime deployment (VPC-peered, dedicated, or regional boundary); Continuous trajectory monitoring, citation-gate tuning, and model fallback policies; Incident response, runbook operations, and versioned agent upgrades
Commercial structure: Implementation scope, recurring operation, included ACU allowance and agreed overage rates.

Explore details

Bring the operational bottleneck and target metric; Forward designs, integrates, evaluates, and ships the production agent workflow.

Automatic Implementation

Executive and operations leaders who want a measured production outcome—not an open-ended consulting engagement or DIY SDK project.

Customer owns: Problem definition, subject-matter edge cases, and historical gold examples; Security review of system connectors and staging environment access; Final acceptance against the agreed baseline benchmark
Forward owns: End-to-end workflow architecture on Agentic OS (Forward HIVE); Connector implementation, custom evaluation rubrics, and Agent RL adaptation when needed; Accepted deployment, operator handoff and agreed post-launch support
Commercial structure: Fixed-scope implementation milestone + transparent ongoing operating fee and ACU envelope.

Explore details

[04 / TRANSPARENT UNIT ECONOMICS]

Metered by Agent Compute Units (`hive_acu`).

Review the applicable ACU conversion schedule, included usage, limits and retry rules before activation. Included model inference must not produce a second Grid token charge.

Inspect all 13 hive.* ACU primitives Turn trajectories into Agent RL

Featured Forward HIVE Agent Compute Unit (hive_acu) meter primitives
Meter Event	Unit	Conversion	What it covers
hive.agent.task.basic	per task	0.20 ACU	Single-turn or lightweight classification, routing, and structured formatting task.
hive.agent.task.standard	per task	0.50 ACU	Standard multi-step agent execution with retrieval, tool calls, and verification.
hive.agent.task.complex	per task	1.50 ACU	Deep research, long-document synthesis, or multi-hop reasoning across large context.
hive.agent.task.Workflow	per workflow	3.00 ACU	End-to-end orchestrated multi-stage workflow with checkpointing and terminal audit.
hive.tool.execution	per call	0.10 ACU	Authenticated connector call, MCP tool invocation, or structured external API action.
hive.sandbox.execution	per second	0.05 ACU	Isolated code interpreter or document-processing sandbox compute time.

Agent creation / Start with intent

From a business request to an accountable agent.

Work with Forward to turn intent into a reviewable specification: the job, owner, tools, limits, tests and deployment boundary.

01
Describe the work
Bring a business request in your own words. Define the result, systems, exceptions and decisions that need approval.
02
Start from an approved pattern
Scope a reusable workflow, connector and evaluation set. Adapt it to your organization and operating rules.
03
Bring an existing agent
Review your agent or harness against the same authority, deployment and acceptance requirements.

Creation paths, automation and integrations are qualified for your engagement. Review the specification and representative tests before activation.

Continuous improvement / Controlled learning

Every accepted outcome can inform the next improvement.

Build a loop from business execution to measured improvement. Agent feedback helps propose and score changes. Verified outcomes and independent evaluation determine whether they work.

01
Business task
A defined job and acceptance criteria.
02
Governed execution
Authorized context, tools and action limits.
03
Verified feedback
Outcomes, expert review and calibrated agent judgments.
04
Reward-based training
Authorized signals update an eligible policy.
05
Independent evaluation
Held-out quality, policy, latency and complete cost.
06
Approved release
An accountable owner or pre-authorized policy governs promotion and rollback.

Illustrative learning lifecycle. Execution permission does not grant training permission. Training changes eligible model or policy weights; prompt and tool changes have their own evaluation path. Learning cannot expand agent permissions. Deployment, training and release gates are agreed for each engagement.

Explore Agent Learning Review governance boundaries

Put your business agents on a clear operating plan.

Scope business context, integrations, governance, evaluation and usage with Forward.

Explore agent governance

From a business request to an accountable agent.

Six capabilities to scope for your workflow.

Durable multi-step execution & checkpointed state

Scoped tools, sandboxes & tenant controls

Native human-in-the-loop interrupts & budget soft-pauses

Deterministic citation audits & OpenDetector guardrails

Tenant-isolated vaults, document parsing & artifacts

Unified Agent Compute Unit (ACU) metering & attribution

Inspect an illustrative document workflow.

01. Scoped Intake & Policy Binding

Build with the API or ask Forward to implement and operate.

Agentic OS Runtime API

Managed Agents

Automatic Implementation

Metered by Agent Compute Units (`hive_acu`).

From a business request to an accountable agent.

Describe the work

Start from an approved pattern

Bring an existing agent

Every accepted outcome can inform the next improvement.

Business task

Governed execution

Verified feedback

Reward-based training

Independent evaluation

Approved release

Put your business agents on a clear operating plan.

Request access

Talk to us

Data Protection & Telemetry Preferences

From a business request to an accountable agent.

Six capabilities to scope for your workflow.

Durable multi-step execution & checkpointed state

Scoped tools, sandboxes & tenant controls

Native human-in-the-loop interrupts & budget soft-pauses

Deterministic citation audits & OpenDetector guardrails

Tenant-isolated vaults, document parsing & artifacts

Unified Agent Compute Unit (ACU) metering & attribution

Inspect an illustrative document workflow.

01. Scoped Intake & Policy Binding

Build with the API or ask Forward to implement and operate.

Agentic OS Runtime API

Managed Agents

Automatic Implementation

Metered by Agent Compute Units (hive_acu).

From a business request to an accountable agent.

Describe the work

Start from an approved pattern

Bring an existing agent

Every accepted outcome can inform the next improvement.

Business task

Governed execution

Verified feedback

Reward-based training

Independent evaluation

Approved release

Put your business agents on a clear operating plan.

Metered by Agent Compute Units (`hive_acu`).