Skip to main content
Forge separates who performs work, how it runs, what it may reach, how it is governed, and how it is operated.

Workforce

People and agents are workers with identity, organizational placement, relationships, and access. Agent definitions add models, prompts, skills, and tools. Read Workforce, including Personal agents.

Workflows and runs

Workflows are versioned graphs; runs are durable executions of a published version. Factories, schedules, triggers, ingress, files, outcomes, and human tasks build on that lifecycle. Read Workflows and runs, Outcomes, and Use the Workflow Architect.

Capability

Tools expose callable operations. Skills contribute reusable instructions and tool groupings. Procedures run deterministic code in a sandbox. Capability providers resolve business operations to integrations without exposing vendor credentials to the worker. When a workflow needs a skill, agent, or procedure that does not exist yet, synthesis can create a provisional one for review. Read Tools and skills, Procedures, and Capabilities.

Governance

Roles and grants establish possible authority. Trust Fabric evaluates a concrete action and returns a decision. Human tasks record approvals and requested input. Judges produce quality or policy signals; they do not replace authorization. Read Trust Fabric, including its Judges section.

Knowledge

Knowledge records and synchronized sources provide governed context. Source access and record scope continue to apply after ingestion. Read Knowledge.

Platform and operations

Convex persists platform state and runs backend orchestration. Sandboxes isolate code execution. Observability, cost, audit, and emergency controls make running work inspectable and controllable. Read the system overview and Observability.

Relationship summary

A caller starts a published workflow. Each node resolves a worker and its versioned configuration. The worker uses assigned tools or procedures; capability bindings reach external systems. Access and runtime governance decide whether each action proceeds. Runs, files, decisions, usage, and outcomes become operator evidence.

Design principles

Forge makes several choices that affect how you build and operate on the platform.

Persist intent before effects

Durable operations record intent and state around external effects. A timed-out write is still an unknown outcome: retain its idempotency key and reconcile state before another logical intent. Read Idempotency.

Separate possible authority from runtime decisions

Roles and scoped grants establish possible authority. Token scopes only narrow it. Runtime governance evaluates the concrete action and can still refuse or hold it. Read Credentials and access.

Keep secrets behind capability boundaries

Agents and procedures request operations; providers resolve integrations and credentials. Sandboxes receive short-lived procedure sessions, not vendor secrets. Read Capabilities.

Version executable definitions

Agents, workflows, procedures, skills, and tools have lifecycle and version semantics. Publishing is deliberate, and editing a draft does not rewrite an in-flight run.

Treat human interaction as durable work

An approval or request for input has identity, eligibility, state, and history. Design for expiry, rejection, reassignment, and no response. Read Approvals.

Prefer explicit refusal to partial success

Sensitive operations fail closed when authority or provenance is missing. Bounded reads disclose paging or truncation. Branch on structured errors and metadata; do not infer completeness.

Keep interfaces distinct

Console is for people, API v3 for deterministic integrations, and MCP for model-usable tasks and searchable operations. Read Choose an interface.

Evidence does not remove judgment

Forge can record what ran, who decided, and what it cost. Evidence may still be bounded or insufficient, and an allowed action may have a bad outcome. Operational review remains necessary. Use the glossary for exact vocabulary and the architecture overview for system boundaries.