Why we dogfood
The fastest way to make a SaaS product worse is to sell it to customers before you trust it yourself. We do the opposite: every agentic feature we sell runs on our own agency first, often for months, before any client sees the polished version.
That means:
- Our SEO is run by the same SEO agents we sell.
- Our blog content (this post, even) goes through the same verifier we sell.
- Our outbound sequences use the same sequence engine.
- Our lead-gen, our reporting, our brand-system audits — all on the same platform.
When we break something, we break it on us. The bugs get caught here.
What we run today
Memory layer. Every client we work with has a brain — meeting notes, contracts, contacts, Drive folders, Canva assets, competitor research, brand voice — all in the same database, queryable by agents at prompt time. We have ~60 memory files across 20 clients with usefulness + quality scores on each. Low-scoring memories get auto-archived weekly so the brain stays sharp.
Skill catalog. 16 active skills (SEO, ads, content, web-dev, creative, leads, analytics, plus sub-skills). Each skill links to the agents that implement it and the best practices that ground it. ~250 best-practice rows tagged by skill × industry — the verifier reads these before scoring.
Verifier agent. Every memory and recommendation goes through quality + usefulness scoring against (a) industry best practices, (b) the client's contract objectives. If a learning doesn't help us deliver on what we promised, it doesn't belong in the brain.
Competitor research. Per-client competitor tracking with three research kinds: agent-summary (LLM analysis based on what we know), site-crawl (live fetch + extract + analysis), ad-library (Meta + Google Ads Transparency check-ins). Findings surface in each client's brain dashboard.
Live agent monitor. Real-time feed of every agent run across the workspace — provider, model, latency, cost, status. We watch this for anomalies the way an SRE watches request latency.
Smart routing. Tiered AI router: cheap models (Groq, Mistral, DeepSeek) for batch and classification work, Claude for complex reasoning, optional Ollama on a desktop for private inference. The router picks per-call based on the task.
Numbers
These are real, from the last 60 days on our prod system:
- Best practices captured: 254 (21 seeded, 233 LLM-generated, all reviewed)
- Memories scored: the verifier rates the top tier at 0.9-1.0 usefulness against contract objectives
- Average verifier latency per memory: ~4 seconds via DeepSeek
- Agent runs per day across all clients: climbing past 200 as we wire more agents to live data
- Dollar cost per agent run: ~$0.0008 (sub-tenth-of-a-cent on average, weighted by tier)
These are the kinds of numbers you can only have when the platform is the operating system, not a side tool.
What we're still iterating on
Honesty: not everything works yet.
- Cross-client industry rollups. We have per-client brains but not yet a clean way to surface "this play worked for Hurtado, apply surgically to Bar-A-BBQ." The RBAC constraints make it harder than it sounds. Working on it.
- Live ad-library scraping. Our Meta App is still in dev mode, so the ad-library research kind currently builds deep-links for manual research instead of pulling creative directly. Should be unblocked once we get out of dev mode.
- Ollama-on-desktop in prod. We have the Cloudflare Tunnel setup documented but haven't flipped the switch yet. When we do, low-tier inference moves off cloud providers entirely.
- Verifier calibration. The current scoring is well-grounded but tends to cluster high (most ANKR memories score 0.85+). We need a calibration pass to make the gradient meaningful at the top of the curve.
What you can take from this
If you're evaluating an agentic agency, ask three questions:
1. Do they run it on themselves? If their own marketing is a normal-agency operation, the version they sell you will have rough edges they've never felt. 2. Can they show you the audit trail? Every agent run should be visible — what context it had, what it produced, how much it cost. 3. Where do they say no? An agency that says "agents can do everything" is selling vapor. The honest answer involves a clear list of where agents win and where humans still belong.
We say no a lot. The list is in the post above this one.
How to start
If you want to see the system in action — including the parts we're still working on — we do live walkthroughs of the platform for prospective clients. No pre-sales pitch deck, just the actual screens.