Designed and built a multi-agent SI content system with clear controls.
I turned a content workflow into a production system that runs across separate machines. One SI agent plans the work and handles problems. Another agent makes the videos. Safety rules keep people in control when needed. Trusted work can also run on its own.
Making more content was easy. Keeping good judgment was harder.
When I took over content operations for a private ecommerce group, several brand Instagram accounts had no posts. The team needed a simple way to make short videos for each brand. It also needed control over the words, images, costs, and final posts.
The challenge was not just to “make a video with SI.” The goal was one safe system for research, creation, review, checks, drafts, publishing, and learning.
Automate the work between key decisions.
The steps below show the safer mode. Gold steps need a person to approve them. Green steps run on their own. These review steps can be changed or removed for trusted work.
- 01Research
Study formats that worked and learn why people watched them.
- 02Concepts
Turn the research or a team brief into ideas for one brand.
- 03Approve words
A human approves the script. Approved language becomes the source of truth.
- 04Approve frames
Review low-cost images before making the full video.
- 05Generate
Make the video from the approved script and visual plan.
- 06Verify
Check the audio, approved words, and images for known problems.
- 07Approve draft
In governed mode, the finished draft waits for a human yes before publishing.
- 08Publish & learn
Publish approved work and use the results to improve the next post.
The team can start with its own idea.
I added a Creative Hub for ideas from the team. A person picks the brand and video type, then writes a few notes. The system can make one close match or three different versions of the idea.
The team’s brief is the main source. It takes priority over research when they conflict. Brand and safety rules still apply. The idea then follows the same script, storyboard, production, review, and publishing steps. This keeps the human idea easy to trace.
One agent plans. Another agent produces.
The system uses two Claude Code setups on separate machines. The main agent keeps the plan, watches the system, and handles problems. The production agent claims approved jobs and makes the videos. A Hermes agent is also connected to the shared network and passed its tests, but it is not doing live production work yet. More workers can be added later through the same task, budget, status, and recovery rules.
One place to watch the system
The team can see messages, approvals, machine work, task owners, and schedules.
Plans, checks, and problems
Owns the system plan, watches research, sorts issues, and keeps problems away from video work.
Handoffs · questions · replies · status · approvals · saved task state
Live updates + backup checksMakes the video
Claims approved work, makes the video, checks the words and frames, adds captions, creates a cover, and prepares the draft.
Review-ready video + production report
Finished work returns through the same channel, so the team can always find it.
One owner per job
Task claims and locks stop two machines from doing the same job.
Work can recover
Regular status checks show stalled work and help the system restart it.
People keep control
Agents can ask for a choice, but protected actions stay with a person in safe mode.
Room to grow
A Hermes agent is connected and tested. It is ready for a future job but is not doing live production work.
Tested with real production limits: every recorded finished job stayed within its set budget. The system counts costs when a job starts, including any new renders. It stops before work would go over the limit. The team can change the limit for different videos or campaigns.
Safety checks are part of the system.
The system follows clear rules. It does not assume the SI will always make the right choice.
Make it cheap to say no
Review scripts and low-cost stills before expensive production work begins.
Lock approved work
Later steps cannot change words or images after a person approves them.
Stop when a check fails
If a required check fails, the job stops and reports the problem.
People approve by default
In safe mode, the system can make a draft but cannot approve or publish its own work.
Fix small defects only
The system can fix small visual problems. It stops if the fix would change approved work.
Adjust the level of control
Review steps can stay, change, or be removed for trusted work.
One working system, not a pile of SI tools.
I designed and built the full workflow so it would make sense to the people running it.
Workflow and idea intake
Mapped each step, choice, approval, and failure. I also built the Creative Hub for ideas from the team.
Multi-agent setup
Gave each machine a clear job and connected handoffs, task claims, status, and recovery.
Team control view
Created one view for agent messages, approvals, machine work, task owners, and schedules.
Quality and safety
Added rules for approved words, audio checks, image checks, budgets, and publishing rights.
People keep control while the system does the work.
This case study shows the system and my role. It does not reveal the client, brands, products, private tools, code, media, or business data.
SI systems need clear owners, safety checks, and a way to recover.
I can help turn an SI test into a working system. I give each agent a clear job, make progress easy to see, add checks, and choose the right level of human control.
Or email me instead.