How we built 23 agents in 90 days
for an 8-person performance agency.

Strivve's 8-person team managed 14 active DTC brands, each with 60+ active ad sets across Meta and TikTok. Three buyers spent every Monday rebuilding daily P&Ls in Sheets, every Tuesday rotating creative because Meta had de-prioritized last week's top performers, and every Wednesday explaining to clients why CAC was up 12%.

The compounding problem: their creative team produced 40 new variants a week, but only ~12 ever got tested because there wasn't bandwidth to launch them properly. The other 28 sat in Frame.io. Meanwhile their best-performing audiences were fatiguing 3 days before anyone noticed.

They were one client win from a hiring round they couldn't afford.

We started with a 2-week paid Pilot. The deliverable wasn't a slide deck — it was a working agent shipped to production. We picked the highest-ROI pain (creative rotation) and built it end-to-end before quoting the rest.

The architecture sits on Mastra orchestration with a mix of local Llama 4 (Mac Mini M4 in their office, Ollama via cloudflared tunnel) for cheap-and-fast tasks, and Claude Sonnet via Vercel AI Gateway for anything requiring real reasoning. Every agent run streams through a custom OpenTelemetry exporter into Postgres so we — and Strivve — can replay any decision after the fact.

After the Pilot landed, we shipped the remaining 22 agents in 11 weeks across seven swarms: Creative (rotation, fatigue detection, variant generation), Budget (daily reallocation, anomaly stopper), Reporting(P&L stitcher, client-portal builder, weekly Loom narrator), Audience (lookalike refresher, exclusion hygienist), Comms (Discord ops bot, Slack client channels), Compliance (Meta policy pre-checker, brand-voice validator), and Ops (onboarding, calendar, contract).

• 340% ROAS lift over 90 days, measured against the pre-Pilot 28-day baseline. Driven mostly by the creative rotation + fatigue detection agents.
• 38 hours/week of media-buyer time reclaimed across the 3 buyers. They now spend it on client strategy, not Sheets.
• 3.4x more creative variants tested per week. The Creative swarm pulls Frame.io approvals, builds Meta/TikTok-ready cuts, and launches them with budget allocations the Budget swarm sets.
• Zero hiring rounds needed. Strivve added 5 brands without growing headcount.
• $84k MRR managed by the orchestrator, with full audit trails per decision.

The whole thing runs on infrastructure Strivve already had budget for. No enterprise procurement, no vendor lock-in.

We're mid-build on the next swarm — a Forecast agent that projects each client's 30-day revenue from current ad ops state, and a Hiring agent that tells Strivve when their actual capacity (not their headcount) is becoming the bottleneck.

If you're running a similar agency with $50K–$1M/mo in Meta spend and more creative than you can ship, we should talk.