Self-initiated · Agentic systems

Designing an autonomous, 15-agent growth engine

Role
Architect & operator, end to end
Scope
Agentic system design, orchestration, guardrails
Built with
Claude, scheduled agents, shared-state files
Timeline
2026
Audience engagementLayer 1 · 4 agents
Daily engagement sweep Reel comment replies End-of-day second pass Supporter intelligence
Content productionLayer 2 · 5 agents
Trend scouting Comment mining Script writing Idea backlog Rising-creator scouting
Performance feedbackLayer 3 · 2 agents
Per-post scoring (Keep / Watch / Kill) Strategic weekly review
Business developmentLayer 4 · 4 agents
Inbound deal triage Brand prospecting Outbound pitching Media kit & rate card
Command center One daily brief ties all 15 agents together: what shipped, what needs a decision, what to do next.

Fifteen coordinated agents across four layers, sharing state so they stay in sync, all reporting into a single morning brief.

The scoreboard
15 agents
Coordinated across four layers, running on schedule, reporting through one daily brief
~2×
Daily views roughly doubled in the first week of full operation, measured on the platform's own dashboard
20+
Inbound brand deals logged in about a week, including a marquee LVMH (Guerlain) approach
The challenge

Four full-time jobs, every single day.

Running a content business at scale means doing four things well, at daily frequency, all at once: engaging a real audience without it feeling automated, keeping a pipeline full of on-brand ideas, learning from what actually performs instead of guessing, and fielding a constant stream of inbound business without losing deals in an inbox. Done by hand, that is close to a full-time operations team.

I wanted to know whether a coordinated set of autonomous agents could run all four in parallel, on schedule, without ever losing the owner's real voice or judgment on anything that leaves the account. So I designed one, and ran it live.

Replace the ops team with a system. Keep the voice and the judgment human.
The system

Fifteen agents, four layers, one brief.

Each agent runs on its own schedule and reads and writes to a shared set of files, so the whole system stays in sync rather than drifting into fifteen disconnected tools.

  • Engagement keeps the account genuinely present: it works a roster of 100+ accounts, comments in the owner's voice tailored to each post, replies to new commenters, and quietly maintains a self-correcting map of who actually engages back.
  • Content removes the "what do I post today" bottleneck: it scouts still-climbing trends, mines real audience language from comments, and writes complete, ready-to-film scripts every morning, against a living backlog of on-brand concepts.
  • Feedback replaces guessing with evidence: every post is scored at 48 hours and 7 days against the account's own trailing median, verdicts roll up by format, and a weekly review explains why things worked and names next week's moves.
  • Business development keeps money from slipping: it triages inbound PR and paid deals, prospects and tiers brands by real budget, drafts time-sensitive outbound pitches, and keeps the media kit current.
The judgment layer

Autonomy with the brakes designed in.

The hard part of an agentic system is deciding what it may do on its own and what it must hand back to a person. The guardrails are the design.

  • Draft, never auto-send for anything that represents the business: brand replies, rate quotes, and outbound pitches are written for approval, and the system never invents a price or commits to terms.
  • Voice as a contract: every agent is built around explicit, documented brand-voice rules, so output sounds like the owner rather than a generic model.
  • Safety-first pacing: engagement is capped per run and spread across the day, and agents stop immediately on any rate-limit signal, protecting account health over chasing volume.
The outcome, in 8 days

Fifteen agents ran four operations in parallel, unattended, for a week. The numbers moved to match.

The system went live on June 28 and ran on schedule, with a human approving only what left the account. The point was that it held together: engagement, content, feedback, and business development all running at once without drifting or going off-voice. As evidence it worked, the platform's own insights showed the acceleration: in the first seven days it drove 1.6M views, 1.0M accounts reached, and 158K interactions, nothing projected. Daily views ran roughly 114,000 the prior month and about 230,000 in the first full week, close to double, with Reels carrying about 95% throughout. On the business side, the deal agents surfaced and triaged 20+ inbound PR and collaboration opportunities in about a week, including an early-preview approach from Maison Guerlain (LVMH) ahead of an August launch. I treat the reach as proof the autonomous system performed, not as the achievement itself; the achievement is the architecture and the judgment layer that governs it.

Instagram views, last 7 days
Last 7 days1,607,541 views, 1,016,825 accounts reached
Instagram views, last 30 days
Last 30 days4,557,056 views, 2,256,687 accounts reached
Instagram views, last 60 days
Last 60 days7,972,460 views, 3,605,682 accounts reached
Straight from Instagram's account insights, pulled July 5, 2026. Comparing the pre-launch month to the first full week, the daily view rate roughly doubled.
Why it matters

The architecture transfers. The vertical is just where it was proven.

None of the fifteen agents depend on anything specific to the vertical they were proven in. Every content-driven business hits the same four bottlenecks, whether it is a streaming brand's social presence, a fintech's community, a boutique brand, or a creator: someone has to engage the audience, someone has to keep the pipeline full, someone has to turn performance data into a decision, and someone has to keep business development moving.

What transfers directly is the design: the four-layer architecture, the voice-consistency approach, the safety-first pacing on anything touching a live account, and the discipline of drafting rather than auto-sending anything that represents the business. What gets rebuilt per client is the specific voice, roster, formats, and pricing rules. That customization is the actual work.

This is the conviction the rest of my portfolio has been building toward: the highest-leverage design increasingly means architecting the agentic systems that produce outcomes, then designing the judgment that decides what a machine may do on its own.

Looking back

Autonomy is a design material

Building this end to end sharpened how I think about autonomy and trust: the interesting decisions were rarely "can an agent do this," they were "should it, and where does a human stay in the loop." That framing, systems that act plus judgment that governs, is exactly what I want to bring to a product team designing with AI at scale.

Next project

An AI research system for the design team

View