Hours, not minutes
Routine pipelines stretch across tools and take a person to babysit each step.
Autonomous agentic AI platform
Goals in. Production out.
Describe the outcome. Forge plans it, runs it across every tool, and hands back validated, production-ready work — for Art, Tech Art, Production, QA, Design, Product, and any vertical. Machine speed. Human control.
● Live in production✓ Human approvals✓ Full audit trail
Built in-house at Zynga
Built in-house at ZyngaProduction goalPrepare validated character assets for build
Autonomous execution
✓Jira request understood
✓Fabric generated plan
✓Assets imported into Unity
✓Validation suite passed
✓PR created · Jira updated
Production runExecuting…
The leadership case
AI that ships work.
The constraint is no longer ideas. It is the distance between intent and validated production output.
Forge closes that distance. It coordinates specialist agents, models, tools, knowledge, and approvals as one accountable production system—not another assistant waiting for prompts.
The opportunity
Great games are built on creativity — not repetitive production work.
Moving assets, updating data, validating content, coordinating tools, running the same workflows again and again. It's essential — and it quietly consumes the time our teams could spend creating.
Tool sprawl
Photoshop, Unity, Jira, Perforce, Git, cloud — context-switching every run.
Knowledge walks
How-to lives in people's heads; consistency drifts as teams change.
Proven impact
Make the outcome impossible to miss.
Faster execution is only useful when quality, governance, and operational visibility arrive with it.
85 minutes
30 seconds
170× faster
From production request to validated output.
3
Autonomous core agents
Backed by specialized minions.
100%
Traceable execution
Every decision, tool call, and approval recorded.
3Autonomous core agents · +minions
30sPer-workflow execution · from hours
1,112Automated tests green
100%Audited runs · every action explainable
Autonomous execution
Watch intent become output.
Forge continuously plans, acts, validates, and escalates—moving work forward while preserving human authority.
GoalIntent received
ReasonContext gathered
PlanSteps composed
SelectAgents assigned
ExecuteTools operated
ValidateQuality checked
ApproveHuman gate
DeliverSystems updated
Understanding the production goal
Live production run replay
Evidence, not promises.
Replay every decision and action exactly as it happened—from request to production-ready result.
forge-run-1842.logPRODUCTION
✓
Jira request detected
FORGE-248 · Prepare character assets
✓
Fabric plan approved
4 actions · 2 tools · 1 validation gate
✓
Unity agent selected
Matched capability and project context
✓
Assets imported
Materials mapped · prefabs generated
✓
Validation passed
Naming · topology · references · build
✓
Production loop closed
PR created · Jira updated · owner notified
✓ Run completed in 27.0 seconds
Where it stands today
In production, not in theory.
Real agents running real production work today — core agents backed by specialized minions, on one governed runtime.
Technical Art
● 3 core agents · fully autonomous✓Ability-icon pipeline
✓Offers & Dynamic Dialogue
✓DD Iconic Item
PhotoshopUnityJiraPerforceGitS3
Core agents & minions
● Orchestrated autonomyEach goal is owned by a core agent that plans, governs and reports — dispatching minions, focused sub-agents that each master one job, and composing their work into a validated result.
↳Asset & source prep — Perforce · Photoshop
↳Unity import & scene setup
↳Data & config synthesis
↳Validation & vision QA
↳Publishing — Git · Jira · S3
👁 Powered by Fabric Visual Intelligence
Forge doesn't just read — it sees.
Game production is visual — so Forge looks at the work. The Vision Engine analyzes rendered artwork, UI screens, icons and portraits, raising confidence and catching issues before assets reach the game. And because vision runs locally, unreleased artwork never leaves Zynga.
Visual inspectionAnalyzes rendered assets — not just filenames.
Confidence scoringVision feeds execution confidence directly.
Production validationDetects visual issues before publishing.
One production system
Forge, Fabric, and Studio—clearly.
Three distinct layers turn intent into governed execution without hiding who is responsible for what.
Forge
The command layer where leaders set goals, teams monitor outcomes, and humans approve consequential actions.
Goals & orchestrationRun visibilityApprovals & audit
Fabric
The reasoning layer that understands context, composes plans, routes models, and learns from production knowledge.
Planning & reasoningModel routingKnowledge grounding
Studio
The specialist-agent layer that operates production tools and systems to create, validate, and deliver work.
Production agentsTool integrationsValidation harnesses
Human governance surrounds every layer: policy, permissions, approvals, observability, and a complete audit trail.
The platform, in depth
Four parts. One accountable system.
Forge sets the goal, Fabric decides the how, Studio builds the doer, and Weave guarantees the run. Each part is real, running, and inspectable.
Forge
Command planeWhere goals go in and accountability stays in the room.
- Goal intake and live orchestration
- Pipeline & run visibility in real time
- Human approval gates on every consequential action
- Complete, replayable audit trail
- Workspaces per production vertical — Tech Art, Design, QA, Localization
Fabric
Agentic AI runtimeUnderstands the work, composes the plan, routes the model.
- Context reasoning & multi-step planning
- Model routing with capability resolution and failover
- Knowledge grounding (RAG) on your own production data
- Decision engine that classifies and gates every action
- Learns from every completed run
S
Studio
Agent workbenchDesign, certify, and commission specialist agents on a visual canvas.
- Visual agent graph — stages, paths, capabilities
- Bounded dry-run pilots before anything ships
- Certification review with quality gates
- One-click commission from pilot to production
- Versioned, with instant rollback
W
Weave
In-app assistantThe screen-aware assistant — ask, build, navigate, and explain, always inside governance.
- Ask anything about your production work
- Builds and edits agents by conversation
- Navigates the platform for you
- Explains every decision and confidence gate
- Sees your current screen for grounded context
Governance is not a feature — it wraps everything.
The same guarantees hold whether an agent is being piloted or running in production.
✓ Dry-run pilots✓ No real writes until commissioned✓ Certification gates✓ Human approval on consequential steps✓ Confidence-gated execution✓ Resumable, idempotent runs✓ Secret-scrubbed diagnostics✓ Private, local-first AI — data never leaves Zynga✓ Enterprise-ready architecture✓ Full audit & replay
The leadership value
Capacity for creative work.
The human creative craft stays human — Forge automates the repetitive, governed work around it, so teams spend more time creating and less time coordinating.
170×
Faster per workflow
From request to validated output.
Thousands
Production hours / yr
Returned to creative work as Forge is adopted.
100%
Governed & audited
Every decision and action explainable.
0
Data leaves Zynga
Private, local-first AI on our own hardware.
Where it's going
Built to expand across every discipline.
The same runtime, the same governance — new verticals come online as agents, not as new tools to learn.
Automated regression & test authoring.
On the roadmapPipeline coordination & release prep.
On the roadmapConfig, tuning & live-content ops.
On the roadmapAsset prep, variants & validation.
On the roadmapRepetitive build & integration tasks.
On the roadmapEvents, offers & recurring live work.
On the roadmapMove from AI assistance to autonomous production.
Give Forge a goal. Keep the control. Get validated work back.
"The best way to predict the future is to invent it." — Alan Kay
Aravindh Subramanian
Associate Technical Art Director · Zynga · Architect, Forge
asubramanian@zynga.com