From chat assistant to production system
A chat assistant that can write a script, make an image or generate video is not an agentic production system. A real system has to know the project's current facts, who is allowed to change them, why the next step may begin, which assets a failure affects, and which decisions must wait for a human.
The core principle: agents occupy bounded workstations, the database holds project facts, events connect upstream to downstream, and gates control irreversible actions.
Recommended workstation boundaries
| Workstation | Principal input | Principal output | Permissions it must not hold |
|---|---|---|---|
| Market and story architecture | Audience hypotheses, commercial route | Story promise, season arc, test hypotheses | Approving a release directly |
| Writers' room | Approved story facts | Episode packs, dialogue, cut points | Modifying locked identity assets |
| Shot design | Episode packs, asset versions | Shot records, coverage plans | Changing story facts on its own |
| Generation operations | Compiled task packets | Candidate material, generation records | Retrying with no budget ceiling |
| Sound and editing | Approved material, the sound bible | Timelines, mixes, candidate cuts | Deleting provenance and version records |
| QC | Acceptance rules, candidate versions | Issues, evidence, fault routing | Quietly fixing and self-approving |
| Release | The complete approved package | Platform delivery and rollback records | Bypassing rights or an owner's signature |
Separating workstations is not about running more agents. It makes failure locatable, permissions auditable and responsibility traceable.
Project memory is five classes of object
- Entities. Characters, locations, wardrobe, props, episodes, scenes, beats and shots.
- Versions. Immutable versions of scripts, references, prompts, models, sound, edits and deliverables.
- State. The currently approved facts — what a character is holding, knows, is wearing, and how injured they are.
- Events. Who moved state from A to B, when, why, and which downstream objects it affected.
- Evidence. Previews, scores, issue timecodes, rights documents, human annotations and approval signatures.
Chat can read these objects and propose changes; it cannot be the sole source of truth. Otherwise one truncated context, rewritten prompt or concurrent task produces contradictory "truths."
Compile prompts instead of copying them by hand
A prompt compiler selects the minimum necessary context for a task type: the current shot's story function, the character's identity version, wardrobe and prop state, space and light, action phase, prohibitions, model capability and acceptance criteria. It also records the template version, input hashes and model parameters, so results can be reproduced or compared.
A very long context is not a synonym for stability. Unrelated characters, future plot, conflicting references and expired assets all lower execution quality and raise data and rights risk.
Let events propagate dependencies
When a wardrobe version, a narration duration or a shot state changes, the system should not re-run the whole project. The event graph finds the direct and transitive dependencies, marks them stale, generates new task packets, and protects approvals that are still valid.
Changing a word of narration, for instance, can affect the audio master, subtitle timecodes, three shot durations, a music cue, the mix and the release package in turn. Without a dependency graph, a team hunts for the gaps from memory.
Budget, retries and human approval
Every task holds a capability token: which assets it may read, which candidate areas it may write, its maximum spend, its maximum retries, its expiry and its approval conditions. Generation operations may submit candidates and cannot self-approve. QC may reject and route an issue and cannot modify locked story. Release must read the complete rights and acceptance evidence.
High-risk actions stop explicitly in front of a person: story greenlight, character identity lock, over-budget batches, rights exceptions, picture lock and formal release. A good approval package shows only the evidence needed for the decision, and states what approving, rejecting or returning will each trigger.
Automation maturity
Start with one project. Get shot records, the continuity ledger, generation records and QC reports right before adding scheduling and automated evaluation. The mark of scale is not running more models at once. It is that failure rates, cost per approved second, human waiting and version rollbacks are all measurable, and that the next season can reuse the same production semantics.