While the CRM was calmly talking to leads and booking calls, our day was spent reviewing live agent conversations, tracking model costs, and trying to upgrade n8n.

👀 Monitoring live Inbound agent conversations
We continue to closely analyze real conversations with people:

• collecting cases where the agent’s behavior doesn’t meet expectations
• making targeted prompt improvements based on those cases
• expanding the FAQ to cover edge cases and rare scenarios

💰 AI cost control
Since testing is now running on a broad live audience, we’ve enabled strict cost monitoring:

• using models 5.1 and 5.2 🤖
• tracking whether it’s economically reasonable to keep the Inbound agent running in this mode 📊
• collecting spend statistics to evaluate future scale and usage patterns 📈

🛠 Updating self-hosted n8n: surprises from the “tabular” version
The most interesting part started when we tried to move to the latest stable version of n8n for automations:

• almost a full day went into reading guides and attempting to upgrade our self-hosted image ⏳📖
• along the way, we discovered that one of the key Python nodes had changed in a way that critically doesn’t work for us

Current understanding:

• Python nodes can no longer reference outputs from multiple previous nodes beyond just one 🚫
• a large portion of our workflows was built around this behavior — Python was used very heavily 🧠

If this is not a bug but intended behavior:

• we’ll need to rewrite dozens (possibly close to a hundred) Python nodes
• migrating to n8n v2 becomes a very heavy project 🏗

We plan to clarify this behavior in the official n8n support channel before deciding on the next steps for migration.

If you want less chaos in your leads and need a CRM — feel free to reach out 💬

Today we focused on two things: controlling the bot’s conversations and migrating all workflows to n8n v2.

1. Continuing to “train” the Inbound agent
Small things, but important:

• we keep monitoring real conversations
• collect cases where the agent’s behavior isn’t satisfactory
• based on those cases, we make targeted tweaks to prompts and logic so the agent becomes more predictable and genuinely useful

2. New dashboard for agent conversations
I built a handy dashboard to monitor how the bot is working:

• you can see all users the agent talked to during the current day
• a counter shows how many messages the agent sent to each user
• there’s a direct link to our internal CRM by ID, so you can instantly jump into the right conversation
• if you view it from an account where the bot is installed and authorized, usernames are also visible, so you can quickly open Telegram, read the chat, and evaluate the agent’s performance

In practice, this is now the most convenient way to:

• control exactly what the agent is saying
• spot edge or controversial cases
• collect material for further improvements

3. Full migration to n8n v2
The biggest part of the day went into a complete migration of our automations to n8n v2:

• it became clear that the old approach with Python nodes no longer works — you can’t reference more than one node back
• the simplest solution was to fully switch from Python to JavaScript

As a result:

• I refactored over 70 workflows
• all Python scripts were rewritten in JavaScript (thanks to Claude and ChatGPT for speeding up the conversion)
• we reconfigured and relaunched the updated self-hosted instance

Everything is running smoothly now, with all processes already on v2. We’re on a modern n8n version with access to all the latest features and updates. The migration was inevitable — and now it’s finally behind us.

If you want to automate your lead workflows as well, feel free to reach out — we can build a CRM tailored to your use case 💬

While the CRM handles chats and books calls on its own, we’ve been focusing on the migration and continuing to train our Inbound agent 🤖

n8n v2 Migration: now truly complete
It turned out the previous migration wasn’t fully finished:

• All scripts had been converted to JavaScript
• But native Python nodes still didn’t work

At first, this went unnoticed because critical flows had already moved to JS. However:
• In new scenarios, Python nodes are still useful when we just need to work with data from a single previous node
• We had to dig into the image configuration and Docker environment separately

In the end, we discovered that last time everything ran as “everything works except Python containers.” Today we fixed it and tuned the image ✅ — now both JS and Python nodes work correctly, and the system has stabilized.

🗓 A week with the Inbound agent: mistakes, humor, and lots of guardrails
Over the past week, we collected quite a few real-life cases with our Inbound bot:

• Someone waited for us in a café ☕️ because the agent enthusiastically agreed to offline meetings
• In some cases, the bot started discussing money 💸 and deal terms
• There were situations where leads sent Zoom links and didn’t understand why the “person” didn’t join 👩‍💻

These cases highlight the importance of:

• Strictly limiting what agreements the agent is allowed to make 🚫
• While still keeping the conversation lively and covering a wide range of topics 💬

Working on complex guardrails

• Short and concise ✂️
• As comprehensive as possible 🧩

Fact-checking and personal context
Another topic that came up was personal account factology:

• Leads sometimes know real-life details about a person (kids, personal facts, etc.) and confidently ask about them 👀
• The agent, not having this data in the database, logically “denies” it and says it doesn’t know 🤷

We understand that these facts need to be carefully added to the knowledge base if we want to maintain natural communication 🌱. But there are too many cases to cover at once, so this will be a long and meticulous process: gradually adding context, testing new scenarios, and improving the agent’s behavior step by step 🐢.

If you want a more systematic approach to leads and processes, we can set up the same CRM for you 💻

So far, no issues with the CRM — it handles leads on its own ✅
What we are still working on is educating the Zero Inbound agent 🤖: teaching it not to invite everyone to join the team and not to duplicate replies.

Filtering out “job requests” and cleaning up the prompt 📝

While testing the Zero Inbound agent, a new pattern surfaced:

• many leads reach out about jobs, vacancies, or generic “give me a job” scenarios
• previously, the agent was too agreeable — saying “we’re open to everything” and inviting everyone to a call

We’ve cut that behavior out:

• updated the prompt so the agent doesn’t take on hiring commitments
• simplified and softened responses so they’re lighter and more to the point

💡 Deep refactor of the message pooling logic
Most of the work focused on how message pooling works — when the agent needs to process not a single message, but a batch of messages sent by a lead in a row.

⚡️ Intended design:

• the first incoming message triggers the workflow
• all following messages go into a waiting pool: marked as read, but not triggering anything
• the agent waits for the pool to fill and then sends one response that accounts for the entire message sequence

What went wrong 🧐:

• in reality, two independent flows were running
• the agent couldn’t “see” its own previous replies within this logic
• as a result, it could respond to the second or third message with essentially the same reply again

What we changed ✅:

• introduced a full lock on parallel processes until the agent finishes generating a complete response (30–90 seconds of work)
• updated the logic so the agent’s own replies are also considered within the flow

⚙️ Result:

• the probability of duplicate replies is now minimal — still theoretically possible, but very unlikely
• achieving 100% protection would require a much more complex architecture, which we’ll only move toward if the current version proves insufficient

If you want to bring order to your leads and CRM funnels, feel free to reach out 💬

Today the focus was on analytics and making it easier to monitor how the Zero Inbound agent behaves in real conversations 🤖💬

📊 More flexible dialogue stats
Previously, you could only see basic stats for the current day: which accounts the agent interacted with and how many messages were sent.
Now we’ve built a more flexible system directly into the bot menu — adding a stats button with period selection options:

• Today 📆
• Yesterday ⏮️
• Current calendar week 🗓
• Previous calendar week ⏪

This makes it much easier to track who the agent interacted with and how actively over different periods 👀

⚠️ Current limitations
Right now, stats are calculated from our internal CRM, where accounts using the bot are linked. The number of messages and conversations is pulled from this database.
Because of this, there’s an important limitation:

• If an account isn’t in the CRM 🚫, there’s no source for stats — data simply won’t appear for that account

🔜 Next step we’re planning

• Enable stats to pull data directly from the bot’s own database (Postgres) 🗄, where all messages sent after the bot was connected to an account are already stored.
This will make analytics more universal and independent of whether the account is in the CRM ✅

If you want complete transparency on leads and CRM activity — reach out 💬

Silicon Valley Ai Agents LaunchPad (from the Palo Alto Research Laboratory) is repping $NORAH 🌟

What is $NORAH? 🧐

$NORAH is the first lithium-backed RWA token, offering fractional exposure to revenue linked to a 1,000 t/day lithium concentrate plant in South Western Nigeria, issued by Norah Resources PTY LTD (Australia) with 5 licensed mining locations in Nigeria.

🏦 The Market Gap
• Lithium access is mostly institutional.
• Many RWAs lack real production linkage.
• Investors want industrial, not narrative-only, exposure.

🪙 Why $NORAH
• Asset-backed: tied to lithium concentrate output/off-take.
• Revenue-linked: token represents participation in mining revenues.
• Verifiable: on Toronet (MSB + PCIDSS).
• Diversifying: commodity-cycle exposure.
• Accessible: commodity exposure via wallets.

📈 Why Lithium, Why Now
• EV + storage demand keeps rising.
• Supply remains structurally tight.
• Physical commodity provides an anchor.

📊 Project Snapshot
• Plant: 1,000 t/day (SW Nigeria)
• Issuer: Norah Resources PTY LTD (Australia)
• Licences: 5 locations (Nigeria)
• Docs: whitepaper live

✉ Call to Action
DM us for introductions!
@tonyssd @platinumvc1 @tonydzi

Important Links, Token Metrics, etc:
🌐 Website | 📱 Twitter | 📱 Telegram
📥 [email protected]

About Ai Agents LaunchPad:

Proudly made in San Francisco, USA in Stealth Mode..

our Pad was established thanx to:

PaloAlto Ai Research Lab, Silicon Valley Syndicate Ventures, local VCs & Angels, Serial Founders (with successful exits), engineers from leading tech Co (Google, Meta Etc), along with Stanford alumni

Follow us:
📱 X (Twitter) | 🔝 Telegram

Yesterday we put our CRM and bulk messaging to the test in the real world — sending holiday greetings from all our accounts at once 🎄📨

🎯 The task
We’ve got dozens of accounts, and the relevant leads we want to congratulate are scattered across all of them 🌍
At the same time, our current broadcast system can’t “fire” from any account we choose — each campaign is tightly tied to a single account and its chats 📲

🧩 How we handled it
Instead of one “magic” blast, we built a cascade:

• we moved from smaller accounts to bigger ones 📈
• had to create dozens of separate campaigns 📑
• our AI agent inside the CRM helped a lot — it sped up preparing and configuring all these broadcasts 🤖

Then we launched the holiday campaign and are now actively monitoring how it’s performing 👀📊

🚧 Telegram limits & manual tweaking
Of course, it doesn’t work perfectly in full auto — we hit Telegram’s limits:

• even in “warm” chats, some accounts can’t message users if the last conversation was too long ago 🕰
• because of that, part of the campaign has to be pushed manually — we track who actually got the message and who didn’t ✅❌

Still, the large holiday blast across dozens of accounts went live 🚀
In parallel, we’re collecting insights on how to make this format more automated next time ⚙️

If you want a CRM that actually works, reach out 💬

Although we’re still tweaking the CRM, the main focus right now is on “raising” our Zero Inbound agent — teaching it to behave like a polite sales rep, not an annoying spammer 🤝📩

Yesterday we kept polishing the Zero Inbound agent’s behavior in real conversations and baked a set of important changes into the prompt 🛠

1. Asking about budget — without an interrogation 💸
Previously, the agent could circle back to the budget question several times in a row and came off as pushy.
Now the prompt explicitly says:

• the agent asks about budget only once before suggesting a call, for example:
“By the way, do you have a budget range in mind for tools like this?”
• if the lead doesn’t answer or changes the topic, the agent doesn’t push and never asks about budget again 🚫

2. No Calendly — only @manizha_business 📆➡️👩‍💼
The agent sometimes tried to be proactive with Calendly and attempted to book calls directly in chat.
Now:

• it never suggests Calendly and never books calls itself
• if a lead asks why there’s no Calendly link, the agent explains that the calendar is overbooked
• for anything call-related, the agent always redirects to @manizha_business

3. The correct email ✉️
Before, the agent was giving out a made-up email: [email protected]
Now the correct address is hard-coded into the prompt:
[email protected]

4. Budget — only if we’re talking about the CRM 📊
Sometimes the agent asked about budget completely “out of the blue”, even when the CRM hadn’t been mentioned yet.
We’ve tightened that up:

• the agent never asks about budget if the chat hasn’t mentioned the CRM (our product) before

5. No promises to pay for anything 💳🚫
Previously, the agent could casually say we were ready to pay for someone’s services.
Now there’s a strict guardrail in the prompt:

• whenever payment or money comes up, the agent clearly says it’s an AI agent and can’t handle payments
• for any billing or financial questions, it redirects to @manizha_business

In parallel: the holiday broadcast continues 🎄📨
The holiday messages are still going out, and along the way we found another important CRM-related detail:

• right now, if a user’s account is deleted, we only see an error when trying to message them ⚠️
• but the fact that the account is deactivated isn’t stored anywhere explicitly in the database

The next step is to start marking deactivated users directly in the CRM ✅
So the system understands that this contact is no longer reachable and doesn’t try to send them anything in future campaigns 🔁

If you need a CRM tuned for your lead flow, we can build it around your use case 💬📈

CRM and the holiday outreach did their job, but the day still turned out unexpectedly intense 😅 — the agent pitched a non-existent trading bot, costs suddenly spiked, and Telegram blocked our domain.

1️⃣ Conversation review — new edge case: the agent actually booked a demo with a lead by pitching a solution… that we don’t even have yet 🤷‍♂️
2️⃣ Finished the Christmas outreach 🎄 — huge wave of replies → agent got overloaded → spend increased 💳
3️⃣ Telegram stopped sending webhook events to n8n just because it didn’t like our domain — we switched the domain and everything started working again ⚙️

If you want to pull all leads into one CRM and never lose a conversation again — DM us 💬