Marketing — social content productionMulti-agent orchestration

A four-agent content workforce that took carousel production from six hours to eight minutes

A social team producing 50+ Instagram carousels a week at six hours each — research, writing, three rounds of review — rebuilt as four specialised agents with a coordinator.

A comparison of the same carousel produced two ways, drawn as proportional time bars. The manual path runs three hours of research, two hours of writing and one hour of review — six hours in total — with three or more revision rounds on top when a junior creator wrote it. The agent path runs the same three phases in two, three and three minutes, eight minutes in total, with a 98 percent client approval rate on the first draft. Below, three outcomes: 420 thousand dollars saved annually against 480 thousand of prior spend, 300 percent more content from the same three people, and 150 pieces a week up from 50-plus carousels.
Same three phases, same team, same brand voice. Only the time changed — and then the volume.

Project snapshot

Client
A social media team producing 50+ Instagram carousels weekly with a three-person core
Industry
Marketing and content production
Business function
Social content research, writing and review
Challenge
Each carousel took 6+ hours across research, writing and review, costing $480K annually. Trending topics were missed because research cycles were too slow, junior creators needed 3+ revision rounds, and senior staff were burning out on repetitive work.
Solution
Four specialised agents: a Retriever that researches topics across the internet and an internal knowledge base, a Writer with explicit thinking capability, a Reviewer running self-correction loops, and a Coordinator orchestrating the workflow. Memory systems learn brand voice over time.
Result
Carousel production dropped from 6 hours to 8 minutes, with 98% first-draft approval, $420K in annual savings and a 300% increase in content volume from the same team.

Key outcomes

8 min
Per carousel, down from 6 hours
98%
Approval rate on first draft
$420K
Annual saving in production cost
300%
More content, same three people

The client

A three-person team on a fifty-carousel week

The social media team was producing 50+ Instagram carousels every week. Each one took over six hours end to end: research, then writing, then review.

Three people, fifty pieces, six hours each. The arithmetic does not work, and it showed.

The challenge

The cost was never really the money

$480K a year went to content creation alone, and roughly $23K a week evaporated in lost productivity. But the spend was the symptom.

The manual path

6+ hrs
per carousel across research, writing and review, at 50+ carousels a week
3+
revision rounds whenever a junior creator wrote the piece — senior time spent twice

The real damage was compounding. Trending topics were missed because research cycles could not keep pace with the trend. Junior creators needed three or more revision rounds, so every piece they touched consumed senior time twice over. Senior staff burned out on work they were overqualified for. And quality was inconsistent across team members, because six hours of judgment is six hours of someone’s judgment.

Constraints

What we had to design around

  • VoiceBrand voice is the product. A faster pipeline producing off-voice content is worse than a slow one, so quality could not drop at all — not “not much”.
  • VolumeOutput was already the bottleneck, so the system had to absorb more work rather than merely do existing work faster.
  • ReviewJunior-to-senior handoff was the expensive step. Fixing generation without fixing review would have moved the bottleneck, not removed it.
  • FreshnessResearch had to reach live sources, because the whole problem was topics going stale before publication.

Our approach

Four roles, not one assistant

A single model asked to research, write and edit does all three at the level of its weakest. And it does them in sequence inside one context, so a weak research pass quietly poisons the writing without anyone being able to point at where.

The team’s own process already had four distinct roles in it. We built four agents to match, which meant each one could be measured, tuned and replaced on its own.

The solution

A content workforce with a coordinator

  1. Retriever Agent — researches topics across the open internet and an internal knowledge base, collapsing the phase that used to gate everything downstream.
  2. Writer Agent — generates carousel content with an explicit thinking step rather than one-shot generation.
  3. Reviewer Agent — polishes through self-correction loops, absorbing the revision rounds that used to land on senior staff.
  4. Coordinator Agent — orchestrates the workflow, so the handoffs become the system’s problem instead of a person’s.
A brief enters the Coordinator Agent, which orchestrates the workflow so handoffs are the system's problem rather than a person's. Three specialists run in sequence beneath it: a Retriever that researches the topic across the open web and an internal knowledge base, a Writer that generates the carousel with an explicit thinking step rather than one-shot generation, and a Reviewer whose self-correction loops absorb the revision rounds that used to reach senior staff. A brand-voice memory layer spans all three and learns over time, which is what holds quality steady while volume triples. The output is a finished carousel approved on the first draft 98 percent of the time.
The coordinator owns the handoffs. That is the part a person used to do, badly, fifty times a week.

Underneath: GPT-4 powered agents with custom thinking loops, memory systems that learn brand voice over time, multi-tool integration across web search, knowledge base and feedback loops, and real-time orchestration between the specialists.

2 minresearch phase, against three hours manually — the step that used to gate every piece behind the news cycle

Responsible by design

The memory layer is what protects the voice

Speed is the easy half of this problem, and the dangerous half. A content pipeline that triples output while drifting off-brand does more damage than the bottleneck it replaced, because the drift is invisible until a campaign lands wrong.

So brand voice is not held in a prompt. It is held in a memory layer that learns from what the team approves and corrects, and it sits underneath all four agents rather than being restated in each one. The Reviewer’s self-correction loop reads from the same place.

The 98% first-draft approval rate is the measurement that matters here — not because it is high, but because it is the number that would fall first if the voice started slipping.

Results

Measured on a live carousel, end to end

  • A finished carousel in 8 minutes, against 6 hours: research 2 minutes against three hours, generation 3 minutes against two hours, review and polish 3 minutes against one hour.
  • 98% client approval on the first draft, against a prior process where junior creators needed 3+ revision rounds before a piece was shippable.
  • $420K saved annually, against $480K of annual spend on content creation alone.
  • 300% more content from the same three people— from 50+ carousels a week to 150 pieces a week, with no measurable quality drop.

Beyond the numbers

What else changed

The client now produces 150 pieces of social content weekly with the same three-person team. The seniors stopped being a revision bottleneck and went back to creative direction, which is what they were hired for.

8 minfrom brief to a carousel the client approves without a revision round

The second-order effect is the one the team noticed: because research now takes two minutes, a topic can be picked up while it is still a topic.

A note on the arithmetic: six hours to eight minutes is a 97.8% reduction, and the engagement reported it as 95%. The per-phase breakdown above is the more defensible framing, and it is the one we would stand behind.

If your content team is spending its senior hours on revisions

The fix is usually not a better writing model. It is separating the roles your process already has, so each one can be measured on its own.