Activity
Mon
Wed
Fri
Sun
Nov
Dec
Jan
Feb
Mar
Apr
May
Jun
Jul
Aug
Sep
What is this?
Less
More
146 contributions to Brendan's AI Community
Do we actually need an AI agent for this?
One thing I keep noticing in AI automation: People sometimes use an AI agent where a simple workflow would be better. For example: Form submitted → validate data → update CRM → send confirmation That probably doesn’t need an autonomous agent. But: Customer asks a complex question → search knowledge base → understand intent → decide what information is needed → use tools → respond → escalate if necessary Now an agent starts making more sense. So I’m curious: **Where do you personally draw the line between a workflow and an AI agent?** Would love to hear how other builders are deciding this in real projects. ---
1 like • 10d
I usually draw the line at decision-making. If the path is predictable, use a workflow. If the system needs to interpret context, choose actions, and adapt dynamically, that’s where an agent adds value.
What happens when your AI agent fails in production?
I’ve been thinking about this a lot lately. Building an AI agent that works in a demo is one thing. Building an AI agent that keeps working when: - the API times out - the LLM gives an unexpected response - a webhook fails - a CRM is unavailable - a customer gives messy input - an automation runs twice - a tool returns incomplete data That’s a completely different challenge. So I’m curious: How are you handling failures in your production AI workflows? Do you have: - retries and fallbacks? - structured error logging? - human handoff? - execution monitoring? - automatic alerts? - rollback or recovery logic? I’m increasingly convinced that reliable AI automation is less about making the agent smarter and more about designing the system around failure. What does your production setup look like?
2 likes • 12d
100% agree. The real challenge starts when things go wrong in production. Retries, validation, clear fallbacks, and human handoff have been really important for making AI workflows more reliable.
From Manual Bookings to a Fully Automated Hospitality Platform — A Case Study
Over the past several weeks, I've been building and hardening a complete booking automation system for a luxury vacation rental — Intervale Mountain Lodge, a 10-bedroom mountain retreat in New Hampshire. Sharing the full scope because I think the unglamorous parts are the actual story. What it replaced: Manual phone screening, hand-calculated quotes, texted door codes one guest at a time, no real visibility into what dates were actually available. What's live now: 🔹 Real-time pricing engine (Supabase) — server-side calculation of nightly rates, fees, and taxes, returned as a locked, tamper-proof quote 🔹 Airbnb-style admin pricing calendar — the owner can set per-date rates, block dates, and adjust global defaults from a password-protected dashboard, syncing instantly to the guest-facing site 🔹 Two-way Airbnb calendar sync to eliminate double bookings 🔹 Multi-step guest qualification form, feeding a structured inquiry straight into the CRM (GoHighLevel) 🔹 Password-protected guest portal — door codes, WiFi, house rules — served securely per booking, with rate-limited login and unique per-booking passwords 🔹 Full review system — guest submission funnel, admin moderation (approve/reject/archive), and a public reviews page blending direct guest reviews with synced Airbnb reviews 🔹 Security hardening pass — Row-Level Security on every table, admin-authenticated endpoints, credential rotation, and a full frontend secrets audit 🔹 Cross-team integration debugging — resolved multiple production issues surfaced by the automation team's live n8n workflows: a 500 error from an unhandled edge case, a field-naming mismatch between systems, and a data-mapping gap between the website and CRM contact records 🔹 Architecture documentation — when platform-stability concerns came up around external integrations vs. native CRM features, documented exactly which parts of the system are external, why, and what the actual upgrade risk is (spoiler: one integration point, clearly isolated)
1 like • 20d
This is a great example of where the real value in automation comes from. The AI-assisted building is impressive, but making the whole system reliable, secure, and actually usable in production is the hard part. Especially liked the focus on cross-system debugging and documenting the architecture instead of just patching things until they work. That’s what turns a demo into a real business system.
Daily build log — Intervale Mountain Lodge 🏔️
Big shift today — not a bug fix, an architecture change. Our partner shipped a full guest review + admin moderation system directly on the website, which meant tearing out a chunk of what we'd already built and rebuilding it around the new design. What changed: The site now has a /admin/reviews panel where guests submit reviews and Steve approves/rejects them, posting straight to Supabase. That whole loop — collect rating, match to booking, route to Steve for approval, publish — now lives on the website side, bypassing GHL/n8n entirely for that part. That means our original SMS-based approach (text guest for a rating → parse reply → match booking → text Steve a one-tap approve/reject) is retired. Workflows 5B, 5C, and 5D are dormant — left in place, but no longer part of the live flow. What I rebuilt: Workflow 5A (the post-checkout trigger) got reworked instead of scrapped. It still fires the day after checkout and still marks the booking so guests don't get pinged twice — but now it sends both SMS and email with a direct link straight to the review page (/review?opportunityId=...), handing the guest off to the partner's new system instead of trying to collect the rating over text. Testing: Ran the whole thing live, start to finish — inquiry → approval → payment → confirmation → 5A firing on schedule → guest gets the SMS link → submits a review on the partner's page → shows up correctly in Steve's Pending Reviews panel with the right rating, comment, and booking reference attached. ✅ Complete and verified end-to-end. Good day to simplify instead of add — fewer moving parts on our side now that the website owns review collection and moderation. Next up: still Calendar Booking Automation (Workflow 3), still waiting on Stripe to go live for a real payment test. Still blocked on: Stripe connection, A2P phone verification, and the custom email sending domain (deliverability fix flagged yesterday) — all outstanding before a full real-data run across the workflows.
1 like • 23d
@Brendan Jowett Appreciate that! I’ve definitely been learning that sometimes the better move is knowing when not to force your original approach. Keeping the existing workflow functional while adapting it to the new requirement gives you flexibility without throwing away the work you’ve already done. And honestly, documenting the “why” is something I’m trying to get better at too.
1 like • 21d
@Mouldi Nouri Yeah, exactly. Shorter declarative sentences make the reasoning much easier to follow without stripping out the useful context. And matching directly on opportunityId is a much cleaner approach—no need to rely on parsing or interpreting a text reply when you already have a reliable identifier. That kind of deterministic logic is what makes the workflow much more robust.
Daily build log — Intervale Mountain Lodge 🏔️
Two bugs today, and both turned out to be "the code's not lying, you're just misreading it" situations. Bug #1: SMS approvals silently failing We've got a workflow where an SMS like "approve [guest name]" matches against open inquiries and moves the booking forward. Testing it kept returning "couldn't find a pending inquiry matching ''" — empty name, every time. Dug into the parsing code and found it: the node strips the command word off the front of the message to isolate the name — so it needs the message shaped exactly like "approve John Smith", command first. Any other phrasing and the regex strip does nothing useful, name comes out blank, and it silently routes to the fallback branch instead of throwing an error. No crash, no red flag — just a wrong-looking result that looks exactly like a real matching bug until you actually read the extraction logic line by line. Bug #2: The invoice email that wasn't missing Got a report that a test guest never got their invoice. Went straight to the API response from the invoice-send call, fully expecting to find a failed request or a wrong recipient field. Instead: both the SMS and email legs came back successful. "msg": "Email queued successfully", correct recipient, correct invoice link, all of it. So where'd it go? Spam folder. The email was sent from GHL's shared sending domain rather than a verified custom one, and Gmail flagged it as "similar to messages identified as spam in the past." Good reminder for myself: before assuming an automation is broken, check whether the thing actually did happen and just landed somewhere unexpected. Would've been easy to burn an hour re-debugging a workflow that was working exactly as built. Real takeaway from today though — this isn't just a one-off test annoyance. Every guest-facing GHL email (invoices, confirmations, review requests) is riding on that same shared domain, so it's a deliverability risk across the whole system, not just this workflow. Adding "set up a verified custom sending domain (SPF/DKIM/DMARC)" to the pre-launch checklist.
2 likes • 24d
@Brendan Jowett Yeah bro, exactly! The raw-value logging is a great addition—I’ll definitely start adding that before the matching logic so these issues are easier to spot. And agreed on the custom sending domain too; deliverability is one of those things that’s easy to overlook until it causes problems in production.
1 like • 21d
@Mouldi Nouri Exactly. That’s the kind of failure that can be easy to miss because technically everything “worked.” Logging those silent fallbacks explicitly makes a huge difference—you want to know when the system is making an unexpected decision, not just when a node throws an error. Empty or missing critical fields should definitely be treated as a visible failure case.
1-10 of 146
Okasha Khan
5
141 points to level up
@okasha-khan-1715
AI & Automation Builder | Helping businesses streamline operations with AI, automation, and intelligent systems.

Active 21h ago
Joined Feb 24, 2026
Powered by