Media AI agent · Live
Highlight-Clipping & Live-Editing AI Agent
Propose goal, incident and session clips from the live feed — each carrying the decision it turns on, its captions and the window it may run in, and a named producer publishes.
- Typical delivery
- 4–6 weeks
- Deployment
- Your stack
- Publishing
- Named producer
- After launch
- Agent Care
What this agent does
Cuts the clip while the event runs
In
- 01
Take the live or just-finished feed with its timecode, the venue event log and the gallery's own flags beside it.
- 02
Read the clip window for that competition and territory, the caption stream and the do-not-clip list in force.
Reason
- 03
Mark candidate moments from the vision, the audio and the event log, and note what preceded each one.
- 04
Set in and out points around the whole passage, so a review, a reversal and the outcome stay inside the clip.
- 05
Carry caption text and the audio bed through the cut, and measure short-form loudness on the clip itself.
Decide
- 06
Hold what the do-not-clip list names: a serious injury, crowd disorder, a pitch invasion, a minor in shot.
- 07
Route a moment still under review, or a window closed in that territory, to the person who owns it.
Out
- 08
Hand the producer the clip with the passage it came from, the decision as it stood and the window it sits in.
- 09
Retain the feed timecode, the event log read, the holds raised, the producer's recut and where the clip ran.
Product statement
The agent proposes clips and holds them. It publishes nothing itself, makes no synthetic commentary voice or altered likeness, and a named producer decides what goes out.
Example workflow
One clip, end to end
- 1Feed runningThe match or session feed, its timecode, the venue event log and the caption stream
- 2Moment markedVision, audio and the event log agree something happened, and the passage around it is read
- 3Clip boundedIn and out points set around the whole passage, with captions and the audio bed carried through the cut
- 4Checks appliedDo-not-clip list, the status of the decision, the clip window for that territory, captions and short-form loudness
No human action required
Stages 1 to 4 run without a person in the loop — the feed is watched and the clip bounded before anyone is asked to look at it. A hold on the passage ends that stretch.
Goes to the producer to publish.
Stops with the person who owns it.
Gallery producer
Reads the clip beside the passage it was cut from, the decision as it stood and the window it may run in.
Publish · Recut · Send for review
Published — handed back
- 6Published by the producerPosted to the channel only after a named producer publishes it; the agent posts nothing itself
- 7Outcome reviewedClips recut before posting or deleted after it, corrections issued, holds raised, by competition and surface
Recuts
Edits the producer makes before publishing are counted in the evaluation.
What should not run autonomously
Human approval stays in control
Outside the boundary — human approval required
8 items
- Sending a clip out to a channel or partner.
- Clearing a clip against a rights or licence window.
- Overriding a hold on injury or distressing material.
- Deciding a minor may appear unblurred.
Automation boundary
Agent acts unaided
- ✓Mark candidate moments from the vision, the audio and the event log.
- ✓Bound the clip on the whole passage, not the moment alone.
- ✓Carry captions and the audio bed through the cut.
- ✓Raise every hold, closed window and unsettled decision it finds.
Write actions run only inside the approval boundaries agreed during implementation. Posting a clip is not one of them.
- Signing off a clip across a reversed decision.
- Issuing a correction or taking a clip down.
- Authorising a synthetic voice or altered likeness.
- Changing the do-not-clip list or the window rules.
Example output
One clip, annotated
Everything the agent proposes is attached to the passage and the feed it came from.
Clip output · single moment
Illustrative example- Moment
- Goal, 61:12 on feed
- Source
- World feed, live
- Window
- Opens at full time
- Handed to
- Producer — held
- Confidence
- 87%
- Held on
- Decision still under review
As received
The feed, the timecode the moment sits on and the window it falls in — nothing here is inferred.
Evidence used
- Event log, goal at 61:12
- Caption stream, live
- Clip window on file
Why it was held
The award was still being checked. A clip cut now shows a goal that may not stand.
What the score decides
Confidence decides how hard the producer re-reads the passage, not whether it may go out.
Action
PublishRecutSend for reviewValue
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 live moment
From the feed and the event log
Keep the passage, not the moment
Bound the clip on what came before and what came after, so a review, a reversal and the outcome stay inside it.
Have the clip ready before the ask
Candidate cuts, their captions and the window they sit in are assembled while the event runs, so the producer reads a clip instead of building one.
Say what the cut leaves out
A decision still being checked, a foul outside the frame, an answer clipped short of its qualifier or a hold on the passage is raised before anything goes out.
Build an evidence trail
Retain the feed timecode, the passage bounds, the holds raised, the window read, the producer's recut and where the clip ran — on both paths.
Integrations
Typical integrations
Five system groups connect to the same agent. Which of them are in scope is decided in discovery.
Live production
- Vision mixer
- EVS replay
- Multiviewer
- gallery talkback
Event & timing data
- Match event feeds
- Venue timing systems
- Scoreboard
- officials' data
Clip rights & windows
- Rights & royalties platform
- Clip terms
- Territory windows
- embargoes
Agent
Highlight clipping & live editing
- Watches the feed
- Bounds the passage
- Holds for the producer
Publishing & captions
- Social publishing
- owned apps
- Caption and subtitle services
- delivery
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 what goes out
Each control wraps the one inside it. A clip clears every layer before a producer can publish it, and a synthetic commentary voice sits outside all six.
L6 · Outermost — last line of defence
Inward → L1 · closest to the model
- L1 – L2
- Decide what may be cut at all
- L3
- Decides what the clip must carry
- L4 – L5
- Keep the send-out with a person and logged
- L6
- Pulls automation back when signals degrade
How Nestack evaluates it
Evaluate the cut and what it left out.
Coverage runs the whole depth of the workflow, and every layer is cut by slice.
Surface — the clip the producer opens
- E1Final-output evaluationDoes the clip represent what actually happened?
- E2Step-level evaluationDid it read the event log, the window and the caption stream?
- E3Tool evaluationDid it pull the right feed and timecode and write to the right surface?
- E4Hold recallWere injury, disorder and minor-in-shot passages held?
- E5Slice evaluationHow does the cut hold across sports, feeds and territories?
- E6Business outcomeHow much was recut before posting, and what was deleted after?
Floor — the clips recut before posting or deleted after it
Failure modes
Where each failure originates in the agent
Seven failure modes plotted against the five stages of the agent lifecycle. The ones that matter most are the ones a deletion cannot take back.
Agent lifecycle
Direction of processing →
01 · Feed & event data
1 mode
- HC-01
Live incident offered as a clip
An injury or a crowd incident marks as action.
Stage gathers
The live feed, timecode, event log and captions
02 · Cutting
2 modes
- HC-02
Cut across a reversed decision
The clip ends before the award was overturned.
- HC-03
Quote cut off its qualifier
The answer is clipped before the condition on it.
Stage bounds
The passage, its in and out points, the audio
03 · Checks
2 modes
- HC-04
Window read from a stale record
The territory's clip terms changed since the sync.
- HC-05
Captions dropped in the cut
Timing shifts and the clip goes out without them.
Stage checks
Holds, the clip window, captions and loudness
04 · Handover / post
1 mode
- HC-06
Hold not read before posting
The reason for the hold does not reach the producer.
Stage presents
The clip the gallery producer reads and posts
05 · Change / Version
1 mode
- HC-07
Silent window-rule regression
A rule change widens what may be cut live.
Stage tracks
Model, prompt, rule and window-record changes
- Sev-1 · the clip misrepresents or exposes someone
- Sev-2 · the clip breaches a window or ships broken
- Sev-3 · clips drift and more is recut
Affected slices
Unsettled moments are where the cut goes wrong
A decision still being checked, a stoppage that turns out to be an injury and an answer clipped short of its qualifier are where a producer recuts and where a clip is deleted after posting. 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 |
|---|---|---|---|---|
| Reviewed and reversed decisions | 6.5% | 4.2× | Review | |
| Injury and stoppage passages | 4.3% | 2.8× | Review | |
| Interview and quote cutdowns | 2.9% | 1.9× | Watch | |
| Routine in-play scoring clips | 1.2% | 0.8× | Normal |
Bar: recut-and-delete rate lift vs. routine scoring clips · scale 0–4.0× · tick marks the 2.0× threshold
2 of 4 slices over threshold
Evidence-linked improvement
Speed is what makes the check get skipped
A clip that arrives late is worth less, and everyone in the chain knows it. The cycle has to make the check cheap enough that nobody has a reason to skip it.
Improvement cycle · five stages
Switchback — the path turns at Improve and returns at Learn
- 01
Detect
Recuts, deletions or corrections cluster on a competition or a feed.
- 02
Diagnose
Traced to the event log, the passage bounds, the hold rule or the window record.
- 03
Improve
The bound, the hold rule or the window record is corrected with your supervising editor, and version-linked.
- 04
Verify
Re-run against held-out passages, including the clips that were deleted after posting.
- 05
Learn
The deleted clip is kept as a worked passage, and the rule it broke is written into the gallery brief.
Learn → Detect
The return edge. Clip windows and competition terms are re-read each cycle — what could be cut live last season may not be this one.
Typical build scope
Twelve workstreams across six weeks
The build scope read against the delivery timeline. Week structure follows the six-week plan — discovery, feed and window records, cutting, evaluation, integration, then supervised live events and handover.
Workstream
Week 1Week 2Week 3Week 4Week 5Week 6
- 01Workflow discovery and automation-boundary definition.Week 1
- 02Live feed, timecode and event-log assessment.Week 1, Week 2
- 03Do-not-clip list and incident hold rules.Week 1, Week 2
- 04Clip windows by competition and territory.Week 2
- 05Moment detection from vision, audio and log.Week 2, Week 3
- 06Passage bounding and reversal handling.Week 3
- 07Caption carry-through and loudness checks.Week 3, Week 4
- 08Producer review and publishing workflow.Week 4
- 09Evaluation suite, slices and regression passages.Week 4, Week 5
- 10Correction, takedown and notification path.Week 5
- 11Publishing, caption and channel integration.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 competition, one feed | ProductionProduction feed integration | AdvancedMulti-sport / multi-market |
|---|---|---|---|
| Introduced at Pilot | |||
| Clip proposals off a live feed | ✓ | ✓ | ✓ |
| Passage bounded on the whole decision | ✓ | ✓ | ✓ |
| Do-not-clip holds and caption carry-through | ✓ | ✓ | ✓ |
| Clip window read before a clip is offered | ✓ | ✓ | ✓ |
| A named producer publishes the clip | ✓ | ✓ | ✓ |
| Baseline evaluation | ✓ | ✓ | ✓ |
| Live feed and event-log integration | ✓ | ✓ | ✓ |
| Introduced at Production | |||
| Publishing and caption-service integration | — | ✓ | ✓ |
| Review workflow and observability | — | ✓ | ✓ |
| Additional feeds, competitions and territories | — | ✓ | ✓ |
| Introduced at Advanced | |||
| Multi-sport and enterprise controls | — | — | ✓ |
| Build price | From $5,000 | From $8,000 | Custom quote |
- Final build price
- Confirmed after discovery based on competitions and feeds in scope, live production and rights-system integrations, clip volume, territory and language coverage, 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
01The feeds you clip from, and the event log behind them
→Live feed, timecode and event-log assessment
Week 102What must never be clipped, and who decides in the moment
→Do-not-clip list and incident hold rules
Week 103Your clip windows by competition, territory and platform
→Clip windows by competition and territory
Week 204How captions reach a clip, and the loudness targets you hold to
→Caption carry-through and loudness checks
Week 305Clips you deleted or corrected, and the reason given
→Evaluation suite, slices and regression passages
Week 406Named producers, and who owns a correction or a takedown
→Producer review workflow and the correction path
Week 507A live event we can run the agent beside
→Publishing integration, then supervised live events
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 both the evaluation work and the first clips a producer publishes.
Phase
W1
W2
W3
W4
W5
W6
- DiscoveryW1
- BuildW2 – W3
- EvaluateW4 – W5
- Pilot & LaunchW5 – W6
Week focus
W1Competitions and feeds in scope; who publishes and who holds
W2Feed, timecode, event-log and clip-window records wired in
W3Moment detection, passage bounding and reversal handling
W4Caption and loudness checks, review workflow and evaluation suite
W5Publishing integration and the first clips out under sign-off
W6Live events running under your producer, then Agent Care starts
- Reading the band
- The do-not-clip list and the window records are in place by the end of week 2, before a single passage is bounded in week 3.
- At the end of W6
- Clips have gone out under a named producer during live events, with the passage, the holds and the window behind each one on file, 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 clipping agent around your gallery and your live feed.
Show us a fixture you covered last month, the feed and event log behind it, and the clips you deleted afterwards. We'll re-cut that event against your own windows and holds, and show you what it stopped, what it got wrong and how late each clip would have been.
- Typical delivery
- 4–6 weeks
- Deployment
- Your live production stack
- Publishing
- Named producer, always
- After launch
- Agent Care available
Nestack Agents · Highlight clipping & live editing
AGT-ME-06 · Agent Care available after launch