Media AI agent · Discovery
Content-Recommendation & Row-Ordering AI Agent
Order rows, shelves and home screens for a viewer from titles actually available in that market, with age-suitability applied, a viewer-facing reason attached and a named editor releasing the shelf.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Editorial control
- Named editor
- After launch
- Agent Care
What this agent does
Proposes the ordering, not the audience
In
- 01
Read the catalogue and the licence window, territory and platform entitlement each title sits inside today.
- 02
Read the profile's maturity setting, its stated preferences and the surface the shelf would appear on.
Reason
- 03
Rank eligible titles for the profile and state which signal moved each one up or down.
- 04
Assemble candidate rows and shelves against the editorial plan the merchandising team already set.
- 05
Choose artwork and copy from the approved asset set and record which asset was picked and why.
Decide
- 06
Drop any title whose window, territory or classification record does not cover that surface.
- 07
Hold a shelf for a named editor where suitability, a contractual placement or a policy flag is involved.
Out
- 08
Send the proposed ordering to the editor with its inputs, its reasons and the viewer-facing explanation.
- 09
Retain the candidate set, the ranking reason, the editor's change and the shelf as it was released.
Product statement
The agent proposes an ordering and the reason behind it. It does not clear rights, set a classification, or put a shelf in front of a viewer on its own authority.
Example workflow
One shelf, end to end
- 1Shelf requestedA scheduled home-screen refresh, a new row in the editorial plan, or a title going live
- 2Availability resolvedEach candidate set beside its licence window, territory and platform entitlement, stamped with the date the record was read
- 3Ordering proposedEligible titles ranked for the profile, with the signal behind each position and artwork taken from the approved set
- 4Suitability screenClassification and descriptors for that market, the profile's maturity setting, contractual placements and anything withdrawn
No human action required
Stages 1 to 4 run without a person in the loop — the availability read, the ordering and the suitability screen are finished before an editor is asked to look at anything.
Reaches the editor ready to read.
Held with the title named, not shipped.
Merchandising editor and compliance reviewer
A named editor decides what is promoted and clears any suitability question or contractual placement. Nothing reaches a home screen before that.
Approve · Reorder · Send back
Approved — handed back
- 6Released and explainedWritten to the merchandising surface only where the editor released it, with the viewer-facing reason attached
- 7Placement reviewedWhat was pulled or corrected after release, by surface, market and language, and where an availability record had gone stale
Reorders
Every position the editor changes is counted.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Releasing a row or shelf to a live surface.
- Deciding a title's classification or content descriptors.
- Overriding a rights window or territory restriction.
- Changing a profile's maturity setting or kids mode.
Automation boundary
Agent acts unaided
- ✓Order eligible titles into a proposed row for a profile.
- ✓Check window, territory and entitlement before a title is ranked.
- ✓Apply the profile's maturity setting and stated preferences.
- ✓Choose artwork from the approved set and record why it was picked.
Write actions run only inside the approval boundaries agreed during implementation. A named editor releases the shelf.
- Placing a paid or contractual promotion in a row.
- Changing the ranking objective or its weights.
- Reinstating a title withdrawn by compliance.
- Turning off a viewer's stated preference or control.
Example output
One row on a home screen, annotated
Everything the agent proposes is attached to the shelf and the profile it was built for.
Ordering output · single row
Illustrative example- Row requested
- Crime drama, home screen
- Profile
- Teen, UK market
- Availability
- Read today
- Status
- Held for the editor
- Confidence
- 88%
- Reason shown
- Not published to the profile
As received
The row as the editorial plan defines it, the profile it targets, and when availability was read.
Evidence used
- Window open in this market
- Descriptor missing on two
- Teen maturity setting
Why it is held
Two candidates carry no age descriptor for this market, and the surface is a teen profile.
What the score decides
It decides how closely the editor reads the ordering — not whether a title may run.
Action
ApproveReorderSend backValue
Where AI adds value
The same four claims, placed at the point in the workflow where each one applies.
Where the value lands
Value 01 – 04
Every shelf and refresh
From the editorial plan, the schedule and new releases
Carry availability with the title
Candidates arrive with the window, territory and entitlement behind them, read at build time and stamped with the date.
Assemble the shelf before review
Candidate sets, ordering, artwork selection and the reason for each position are finished first, so the editor edits a proposal instead of building one.
Catch the ineligible title early
Titles outside their window or territory, and placements a profile's maturity setting excludes, are held now instead of found in a viewer complaint later.
Build an evidence trail
Retain the candidate set, the availability read, the ranking reason, the artwork chosen, the editor's change and the shelf as released — on both paths.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Catalogue & metadata
- MAM/CMS
- title metadata
- Artwork services
- EIDR identifiers
Rights & availability
- Rights management
- avails feeds
- Windowing and scheduling systems
Profiles & controls
- Profile service
- parental controls
- Preference centre
- age assurance
Agent
Content recommendation & row ordering
- Reads availability
- Orders eligible titles
- Routes for sign-off
Merchandising surfaces
- Home-screen CMS
- row scheduler
- App and TV clients
- search and browse
Observability & evaluation
- OpenTelemetry
- Langfuse
- Supported monitoring/evaluation sources
Integration availability depends on the client's existing systems and API access.
Agent controls
Six layers between the model and the home screen
Each control wraps the one inside it. An ordering clears every layer before a shelf is released, and what a viewer then chooses to watch sits outside all six.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Decide whether a title may be shown at all
- L3
- Decides who clears what is promoted
- L4 – L5
- Keep the reason visible and the trail intact
- L6
- Pulls automation back when signals degrade
How Nestack evaluates it
Evaluate the ordering and what it was allowed to contain.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the row a viewer sees on the home screen
- E1Final-output evaluationWas each title in the row available, suitable and correctly labelled?
- E2Step-level evaluationDid it read the current window, territory and profile setting before ranking?
- E3Tool evaluationDid it write the released ordering to the right surface and profile?
- E4Explanation qualityDoes the stated reason match the signal that actually moved the title?
- E5Slice evaluationHow does performance change across markets, languages and profile types?
- E6Business outcomeHow many placements were pulled or corrected at review, and why?
Floor — the placements pulled or corrected after release
Failure modes
Where each failure originates in the agent
Seven failure modes plotted against the five stages of the agent lifecycle.
Agent lifecycle
Direction of processing →
01 · Availability
2 modes
- RE-01
Expired window surfaced
A row keeps a title the licence no longer covers.
- RE-02
Territory mismatch
A title available in one market appears in another.
Stage gathers
Catalogue, windows, territory and platform entitlement
02 · Suitability
1 mode
- RE-03
Unsuitable title, kids profile
A missing or wrong descriptor passes the maturity gate.
Stage screens
Classification, descriptors and the profile's maturity setting
03 · Ranking
2 modes
- RE-04
New title never surfaces
With no interaction history it never enters a shelf.
- RE-05
Long tail stops appearing
A feedback loop keeps the long tail out of view.
Stage proposes
The ordering, the artwork and the reason for each position
04 · Assembly / write
1 mode
- RE-06
Artwork misreads the title
The chosen image implies content the title does not carry.
Stage writes
Only the ordering a named editor has released
05 · Change / Version
1 mode
- RE-07
Silent ranking regression
A weights or model change moves what is surfaced.
Stage tracks
Model, ranking, artwork-policy and shelf-rule changes
- Sev-1 · a title should not have been shown
- Sev-2 · the row misrepresents what it holds
- Sev-3 · the shelf narrows or degrades
Affected slices
A few cohorts carry most of the pulled placements
Kids and teen profiles, day-one licences and thin-metadata archive titles account for most of the placements an editor pulls, and a whole-catalogue number does not show it. Nestack reports performance by slice, not only in total.
Slice performance — reported separately, not only in aggregate
Illustrative example| Slice | Failure rate | Lift | Lift vs. threshold | Status |
|---|---|---|---|---|
| Kids and teen profiles | 5.3% | 3.2× | Review | |
| Newly licensed titles, day one | 4.1% | 2.5× | Review | |
| Archive titles, thin metadata | 3.1% | 1.9× | Watch | |
| Established, long-running series | 1.2% | 0.7× | Normal |
Bar: pulled-placement rate lift vs. long-running series · scale 0–4.0× · tick marks the 2.0× threshold
2 of 4 slices over threshold
Evidence-linked improvement
A shelf that was right last week may not be right today
Windows close, classifications are re-issued and titles are withdrawn. An ordering nobody re-checks keeps surfacing a title the licence no longer covers.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Pulled placements, wrong-window reports or a narrowing shelf show up in one cohort.
- 02
Diagnose
Traced to the availability read, the suitability record, a ranking signal, the artwork rule or a shelf definition.
- 03
Improve
The record, rule or weighting is changed, re-approved by the named editor and version-linked.
- 04
Verify
Re-run against held-out shelves, including the rows that were pulled after release.
- 05
Learn
The pulled row is kept as a regression shelf and the rule is written into the merchandising standard.
Learn → Detect
The return edge. Every cycle re-reads windows, territories and classifications — a row that was eligible last month is not eligible now.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, catalogue and rights records, ranking and shelves, evaluation, surface integration, then production validation and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Workflow discovery and automation-boundary definition.Week 1
- 02Catalogue, metadata and identifier assessment.Week 1, Week 2
- 03Rights, avails and territory-window mapping.Week 1, Week 2
- 04Classification, descriptors and maturity rules.Week 2
- 05Candidate generation and ranking logic.Week 2, Week 3
- 06Artwork selection from approved asset sets.Week 3
- 07Shelf and row definitions with your editors.Week 3, Week 4
- 08Viewer explanations and preference controls.Week 4
- 09Editor review workflow and sign-off record.Week 4, Week 5
- 10Evaluation suite, slices and regression shelves.Week 4, Week 5
- 11Merchandising-surface integration and rollout.Week 5, Week 6
- 12Observability, deployment and Agent Care handover.Week 6
12 workstreams · 6 weeks · bar shows the weeks a workstream is active — several run in parallel
Final scope and sequence confirmed in discovery
Engagement tiers
What each tier includes
Rows are the capabilities named in each tier's scope. Higher tiers include everything below them.
| Capability✓ in scope · — not at this tier | PilotOne surface, one market | ProductionProduction catalogue integration | AdvancedMulti-market / multi-brand |
|---|---|---|---|
| Introduced at Pilot | |||
| Ranked candidate rows and shelves | ✓ | ✓ | ✓ |
| Availability and window filtering | ✓ | ✓ | ✓ |
| Age-suitability filtering by profile | ✓ | ✓ | ✓ |
| Named-editor sign-off | ✓ | ✓ | ✓ |
| Viewer-facing row explanation | ✓ | ✓ | ✓ |
| Baseline evaluation | ✓ | ✓ | ✓ |
| Artwork selection from approved assets | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Merchandising-surface integration | — | ✓ | ✓ |
| Observability and slice evaluation | — | ✓ | ✓ |
| Multi-market and multi-language catalogues | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-brand and enterprise controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on catalogue size, rights and classification systems in scope, surfaces and markets, profile and control integrations, approval controls and deployment requirements.
- Separate from build
- Build pricing is separate from recurring Agent Care, which covers managed monitoring, evaluations, incidents and verified improvements after launch.
What we need from you
What you bring, and what we build with it
Each input maps to a piece of build scope and a week in the delivery timeline.
You bring
We build with it
01Your catalogue, title metadata and identifiers
→Catalogue, metadata and identifier assessment
Week 102Avails, licence windows and territory records
→Rights, avails and territory-window mapping
Week 103Classifications, descriptors and your maturity rules
→Classification, descriptors and maturity rules
Week 204The shelves and rows your editors already run
→Shelf and row definitions, candidate generation and ranking
Week 305Approved artwork sets and the rules for using them
→Artwork selection from approved asset sets
Week 306Placements you have had to pull, and the reason
→Evaluation suite, slices and regression shelves
Week 407Named editors for what is promoted
→Editor review workflow, then rollout under sign-off
Weeks 5–6
Nothing else is required
Deployment, documentation and Agent Care handover are ours.
Delivery timeline
Four phases across six weeks
Phases are drawn over the weeks they actually occupy. Week 5 carries the first shelves released under an editor's sign-off.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Surfaces and markets in scope; who decides what is promoted
W2Catalogue, avails and classification records wired in
W3Candidate generation, ranking, artwork and shelf definitions
W4Viewer explanations, editor workflow, evaluation suite and slices
W5Surface integration and first shelves released under sign-off
W6Live shelves under editor control, then Agent Care monitoring
- Reading the band
- Avails and classification records land in week 2, before any ordering is ranked in week 3. The bars show that dependency, not a smooth ramp.
- At the end of W6
- Shelves have gone live under a named editor with the availability and suitability record behind every title, then Agent Care takes over monitoring.
- Duration
- Six-week plan shown · typical delivery 4–6 weeks depending on scope confirmed in discovery.
Next step · Media AI agent
Build a recommendation agent around your catalogue and your editors.
Show us the shelves you run today, where your avails and classification records live, and who decides what goes on the home screen. We'll rebuild one live row from your own catalogue and show you the titles it would have surfaced outside their window or above a profile's maturity setting.
- Typical delivery
- 4–6 weeks
- Deployment
- Your catalogue and app stack
- Editorial control
- A named editor releases
- After launch
- Agent Care available
Nestack Agents · Content recommendation & row ordering
AGT-ME-01 · Agent Care available after launch