The problem with stateless agents on a live production
A major US entertainment conglomerate — one whose titles have defined Hollywood blockbusters and global animation franchises for decades — came to us with a problem that looked familiar.
They had deployed a fleet of AI agents across pre-production: a script analysis agent, a visual development agent, a continuity-checking agent, a storyboard generation agent, and a scheduling agent that tried to reconcile all of the above against a $200 million production budget.
Each agent was powerful in isolation. Together, they were chaos.
The script analysis agent would flag a scene change. The continuity agent, running an hour later in a separate session, had no idea the flag existed and cleared the scene as clean. The storyboard agent generated boards for a version of the scene that had already been revised. The scheduling agent, working from its own snapshot of the project state, booked a location that the revised scene no longer required.
Every agent was doing its job. No agent knew what the others had done.
A $40 million lesson in stateless coordination
Before engaging EverNow, the studio had run two prior AI production pilots. Both were abandoned after agents produced conflicting outputs that required weeks of human reconciliation to untangle. The second pilot alone cost an estimated $40 million in wasted pre-production work before the plug was pulled.
The studio's head of production technology described the failure mode plainly:
"The agents are not the bottleneck. The bottleneck is that there is no shared ground truth. Every agent starts its session with a different picture of reality, and there is no mechanism for them to converge."
This is the canonical multi-agent coordination failure. It is not a model quality problem. Claude 3, GPT-4o, and Gemini Ultra all failed in this environment — not because they are not capable, but because capability without shared state is just expensive parallelism.
Designing the EverNow memory layer for a film production
We worked with the studio's engineering team over six weeks to design a memory architecture suited to the realities of a film production. The core challenge was that a Hollywood production is not a software project. Its "state" includes:
- Creative decisions — character motivations, tonal choices, directorial intent
- Technical constraints — location availability, VFX pipeline dependencies, vendor contracts
- Temporal state — which version of which scene is canonical right now, not yesterday
- Human approvals — which changes are greenlit, which are under review, which are blocked
EverNow's memory object model maps naturally to all four categories. Each memory object carries a type, a set of structured attributes, a set of relationships to other objects, a provenance trail (which agent or human created or modified it), and a live sync status.
We defined six primary object types for the production:
| Object Type | Key Attributes |
|---|---|
| Scene | version, status, canonical-flag, director-notes |
| Character | arc-status, approved-traits, voice-direction |
| Location | availability-window, booked-status, VFX-requirements |
| StoryboardFrame | parent-scene, revision, approval-state |
| BudgetLine | scene-ref, committed-amount, forecast |
| AgentDecision | agent-id, timestamp, confidence, human-reviewed |
Every agent reads from and writes to this shared object graph. When the script analysis agent flags a scene, it writes a scene.status = "flagged" mutation to the EverNow layer. Every subsequent agent that touches that scene — in any session, on any model — sees the flag before it acts.
The self-learning loop
The most consequential architectural decision was treating the agent fleet not as a set of independent workers but as a collective that improves its own object graph over time.
Here is how it works in practice:
- An agent completes a task and writes its output as a structured memory object — not as a freeform log, but as a typed, queryable entity.
- A lightweight evaluator agent reviews the output against prior decisions stored in EverNow. If it detects a conflict or an improvement opportunity, it writes a
suggestionobject linked to the relevant scene or character. - Human reviewers see a curated queue of suggestions, not a firehose of raw agent output. They approve, reject, or modify.
- Approvals propagate as relationship updates across the object graph — all agents reading related objects immediately see the updated ground truth.
Over time, the object graph accumulates not just the current state of the production, but the reasoning history behind every significant decision. An agent starting a new session does not just know that a scene was revised — it knows why, what alternatives were considered, and which human signed off.
This is the sandbox becoming self-learning: each production cycle deposits institutional memory that shapes how agents behave in the next cycle. By the third month of the pilot, agents were surfacing relevant precedents from earlier in the production without being prompted — because the relevant memory objects were reachable via relationship traversal in EverNow.
Results from the pilot
The studio ran a full-scale pilot on a mid-budget animated feature, approximately 90 minutes of content across 680 scenes.
Pre-production duration dropped from an estimated 14 months (based on the prior comparable production) to 7.5 months. The reduction came almost entirely from eliminating coordination failures and the human reconciliation work they generated.
Agent conflict rate — defined as two or more agents producing outputs that contradicted a canonical memory object — fell from 34% in the first week to under 3% by week eight, as the object graph accumulated sufficient context for agents to self-correct before writing.
Human review load shifted in composition rather than volume. Reviewers spent less time fixing contradictions and more time making genuine creative decisions surfaced by the suggestion queue. Three directors independently noted that they felt more in control of the AI-assisted process than they had in the prior pilots.
Cost per approved storyboard frame fell 61% compared to the prior pilot. The studio attributed this primarily to the reduction in rework cycles — frames generated against stale context, subsequently discarded.
What changed about how films get made
The studio's production technology lead summarized the shift:
"Before EverNow, AI agents were a powerful tool that required constant human supervision to keep from drifting. After EverNow, they behave more like a disciplined crew — each one aware of what the others have done, and able to flag when something does not fit the picture they share."
This is the transition we believe the entire industry is undergoing: from AI as a collection of individually capable but collectively incoherent tools, to AI as a coordinated system with a shared model of the world it is working in.
Hollywood is an extreme case — high stakes, high complexity, deeply human creative judgment at the center. But the coordination failure the studio experienced is not unique to entertainment. It is the default condition for any organization that deploys multiple AI agents against a shared problem without giving them shared memory.
EverNow is the infrastructure that makes shared memory possible. One change, everywhere updated — whether the change is a line of code, a scene revision, or a character decision in a film that has not been made yet.
Working with EverNow
The studio's deployment is in production as of Q2 2025. We are currently accepting applications from teams in media, entertainment, and adjacent content industries who want to explore similar architectures.
If you are building multi-agent production systems and hitting the coordination wall, request access and we will start with a scoping conversation.