Forge Agents

Less repetitive work. More accountable decisions.

Start with one supervised workflow: support response drafts, warm opportunity qualification, or data validation. Agree the business outcome, test it, and only then consider a controlled launch.

Catalog status: configurable patterns and demonstrations, not plug-and-play customer products. Integrations, customer isolation and security controls must be configured and verified for each pilot. Signal and Foundry have unresolved evaluation findings; regulated workflows require additional review.

10

role patterns

Cloud + local

model routing

Human

approval gates

Pilot first

before live actions

Agent catalog

Choose a starting point. Keep the boundaries yours.

These are configurable operating patterns—not one-size-fits-all bots. Every deployment begins with a scoped workflow and explicit control model.

01 / 10

AI Cost & Operations

Meter

Configurable pattern · pilot validation required

Finds AI spend, quality, and routing opportunities across providers and local models.

  • Usage and cost analysis
  • Quality-gated comparisons
  • Routing recommendations
02 / 10

Chief of Staff

Brief

Configurable pattern · pilot validation required

Converts updates, documents, and decisions into an executive operating brief.

  • Priority synthesis
  • Risk and dependency review
  • Decision-ready next actions
03 / 10

Sales Qualification

Scout

Configurable pattern · pilot validation required

Qualifies warm opportunities and prepares contextual outreach without autonomous sending.

  • Account research
  • Evidence-based fit review
  • Approval-ready outreach drafts
04 / 10

Customer Support

Relay

Configurable pattern · pilot validation required

Drafts grounded answers from approved knowledge and escalates uncertain or sensitive issues.

  • Knowledge-grounded drafts
  • Triage and classification
  • Human escalation
05 / 10

Content Repurposing

Signal

Evaluation findings open · not production-ready

Transforms approved source material into channel-specific drafts while preserving meaning.

  • Source-grounded transformation
  • Multi-channel drafts
  • Brand and claim checks
06 / 10

Document Operations

Intake

Configurable pattern · pilot validation required

Extracts, normalizes, validates, and routes information from documents and forms.

  • Structured extraction
  • Completeness checks
  • Exception routing
07 / 10

Customer Success

Pulse

Configurable pattern · pilot validation required

Prepares weekly account reviews, risks, adoption signals, and retention actions.

  • Health summaries
  • Risk and opportunity flags
  • Weekly call preparation
08 / 10

Agent Builder

Foundry

Evaluation findings open · not production-ready

Designs governed agent packages around a customer workflow, systems, tools, and policies.

  • Workflow specification
  • Tool and approval design
  • Evaluation plan
09 / 10

Administrative Care Preparation

CarePrep

Exploratory · not for clinical use or protected health data

Supports non-clinical health-plan administration with clear boundaries and escalation.

  • Administrative preparation
  • Document and policy checks
  • Mandatory human handoff
10 / 10

Cross-industry Data Automation

Flow

Configurable pattern · pilot validation required

Reconciles operational data and routes exceptions across insurance and other industries.

  • Validation and reconciliation
  • Exception detection
  • Approval-gated system updates

Security architecture

Controls to verify before launch.

Email, calendars, CRM, payments, messaging, files and voice are integration options—not included or connected by default. Each needs authorization, configuration and customer-specific testing.

Deployment design and controls are agreed per pilot.

Draft-only by default

Agents can analyze and prepare work before any external action is enabled.

Explicit approval gates

Email, payment, record changes, and other consequential writes require policy and human authorization.

Tenant isolation

Verify that knowledge, credentials, memory and jobs cannot cross customer boundaries before customer data is admitted.

Encrypted secrets

Require protected credential storage and test that secrets do not enter prompts, logs or public interfaces.

Audit evidence

Verify records for model routes, tool calls, approvals, outputs and exceptions, with appropriate redaction and retention.

Confidence escalation

Low-confidence, regulated, or policy-conflicting work routes to an accountable person.

Customization path

A gated path from pattern to pilot.

01

Scope

Outcome, users, systems, risk, and the work the agent must never perform alone.

02

Connect

Approved knowledge, tools, channels, memory, and cloud or local models.

03

Evaluate

Representative cases, quality gates, adversarial tests, cost, and latency.

04

Operate

Controlled launch, audit review, weekly improvement, and measurable outcomes.

Readiness, not a launch claim

Private evaluations do not establish customer readiness.

The owner console is private and is not a customer trial. Before a paid implementation, the scope must identify the runtime, data boundaries, connected tools, evaluation results and remaining risks.

Capabilities and limitations

Residual Forge / supervised agent pilot

Start with one useful workflow. Prove it under human control.

Check an agent pilot’s fit →

Describe a repetitive workflow, the systems it touches, and what must always require approval.