What we learned shipping agents to our first 100 customers

We crossed 100 deployed Avery NXR customers a little while back. Before that number gets diluted by the larger sample sizes ahead, we wanted to share what the first 100 taught us.

These are the patterns that surprised us most.

Pattern 1: People configure the third template, not the first

We thought the first template a new user configured would predict the most-used template.

Wrong.

What actually happens: users explore. They configure one template, look around, try another, look around, then commit to a third one and build their workflow around it.

The third template configured tends to be the one that sticks. It's the one that matched the user's actual pain after two false starts.

Implication: The platform needs to make the first two configurations cheap (low time investment, easy to walk away). We've simplified the setup flow since learning this.

Pattern 2: Users massively underestimate setup time savings from connectors

A typical user thinks "I'll configure Sophia for meeting follow-ups." They're imagining ~15 minutes.

Then they realize Sophia needs Gmail OAuth (1 click) + their meeting transcript folder (drag and drop) + their calendar (already connected from Anna) + Slack for occasional notifications (already connected).

Time to actually configure: 6 minutes. They'd budgeted 15-30.

Implication: The connector reusability matters way more than we marketed. Each agent after the first is faster to configure because connectors are already established. The marginal cost of agent N+1 is way below agent 1.

Pattern 3: The 80/20 of agent failures is the SAME 3 things

In the first 100 customers, we saw 80% of agent failures fall into one of three buckets:

→ Misconfigured triggers. Agent set to fire every hour when user meant every business hour. Or fire on every inbox message when user meant only messages with specific labels. Easy fixes once spotted.

→ Missing context in prompts. Users wrote prompts assuming the agent knew their context. "Send a friendly follow-up" without specifying tone, length, signature. Fixable with one prompt revision.

→ Untested edge cases in inputs. Agent works fine on normal cases, fails on the weird ones. User didn't think about the weird inputs. Adding a fallback path fixes 90% of these.

Implication: Better defaults + better onboarding catch most of these. We added trigger validation, prompt quality checks, and edge case warnings to the setup flow.

Pattern 4: Users add agents in BATCHES, not steadily

We expected: user installs, configures agent 1, uses it for a while, configures agent 2, uses it for a while...

What actually happens: user installs, configures agent 1, uses it for 1-2 weeks, then on a single afternoon configures 3-5 more agents.

The trigger for the batch is usually a specific event: a stressful week, a free afternoon, a realization that their existing tools aren't keeping up. After the batch, agents run on autopilot for 2-3 months before the next batch.

Implication: Don't pressure users to keep building agents constantly. The natural rhythm is bursty. Make the batching moments productive.

Pattern 5: Custom agents come later than expected

We thought users would start with templates, then build custom agents within 2-3 weeks.

Reality: most users don't build custom agents until month 2 or 3. By then they've forked existing templates several times (which IS customization, just lower-stakes).

The path is: → Week 1: configure template as-is → Week 2-4: fork template, adjust prompts, change triggers → Month 2: take a working forked template, add steps → Month 3: build first agent from blank

Implication: Templates do more than "first impression" duty. They're scaffolding that gets bent slowly toward the user's specific needs over months. Quality of templates compounds for the whole user journey.

Pattern 6: Multi-user adoption follows a single champion

We expected teams to adopt Avery NXR collectively after a decision meeting.

Reality: it's almost always one person who installs Free Desktop, uses it for a few weeks, brings it to teammates one-by-one as their use cases come up, eventually upgrades to Pro tier when team coverage is 5+ people.

The champion model is consistent. No champion = no team adoption.

Implication: Free Desktop matters because it lets champions get started without friction. The "I'll bring this to my team" moment depends on them having undeniable personal proof first.

Pattern 7: Audit ledger gets ignored... until it doesn't

We covered this in [post 164] but it bears repeating from the customer cohort lens.

In the first 4-6 weeks of customer use, the audit ledger is basically unused.

Then, sometime between week 4 and month 3, an event triggers the first ledger query — a customer asks why an AI did something, an exec asks a question, a workflow needs debugging.

After the first query, audit ledger usage becomes regular. Users discover it's the answer to "what did the agent do" questions across many situations.

Implication: The audit ledger has a "discovery moment" that's worth designing for. We've added gentle nudges at week 2 reminding users that the audit ledger exists and showing example queries.

Pattern 8: Cloud-LLM-first users have the hardest transition

The hardest user segment to onboard: people who came to Avery NXR after using cloud-LLM agent platforms (Lindy, Relevance AI) for 6+ months.

Why hard: they have ingrained mental models of how agents work, what configuration looks like, what the cost structure means. The architectural differences (local-first, flat pricing, no token meter) feel like missing features at first.

After 2-3 weeks, the mental model adjusts. After a month, they typically prefer Avery NXR's model. But the first 2 weeks are friction.

Implication: We need migration-specific onboarding content for users coming from cloud-LLM platforms. Explaining what's different and WHY makes the first 2 weeks smoother.

Pattern 9: Excel sync turned into a wedge feature (already covered in [post 167])

Customers who came in for "AI agent platform" and discovered Excel sync midway tend to expand usage faster than customers who came in for one specific template.

The Excel sync use case grows into more agents on the same data → agents on related data → eventually a Avery NXR-centric operational stack.

Implication: We've started featuring Excel sync more prominently in marketing because of the expansion pattern.

Pattern 10: Cost projections are still our biggest competitive advantage

When we lose deals, it's usually to incumbent SaaS subscriptions that "do enough" — not to other agent platforms.

When we win deals, the deciding factor is usually cost predictability at projected scale.

→ "$30K-50K/year if we scaled this on a cloud LLM platform vs $10K/year on Avery NXR" → "We're a 50-person team. The math is obvious."

The cost story does a lot of work for us. Other architectural advantages (local data, audit, no lock-in) are supporting actors.

Implication: Our marketing should keep leading with the cost story. Architectural advantages are real but less viscerally felt than the cost projection at scale.

What we'd tell ourselves in retrospect

→ Make the first two template configurations frictionless. Users explore. Friction = abandonment. → Lean into connector reusability in marketing. It's a hidden value that compounds. → Surface audit ledger earlier. Users don't realize they'll need it until they do. → Build migration content for cloud-LLM-platform refugees. The audience is bigger than we expected. → Lead with cost story. It's the deciding factor more than we initially understood.

What we're doing differently for customers 101-500

Based on these patterns:

→ Simplified template setup (fewer required fields, smarter defaults) → Connector reusability prominently shown in onboarding → Audit ledger nudge at week 2 → Cloud-LLM migration guide → Cost projection calculator on the website

If you're customer 101+, the experience is better than the first 100 had. They're the ones who got us here.

→ avery.software — Free Desktop tier. The first 100 customers' learnings, baked into the product.