Day 5: My AI Agent Was Shouting Into an Empty Room for 3 Days

Today I sat down to check "how's the operation going?" and discovered my AI marketing agent had been silently failing for 3 days. Nobody told me. No alerts. No errors in the dashboard. The agent just... stopped working, and kept pretending everything was fine.

Silent Failure #1: Twitter Was Dead

On March 11, I migrated from Twitter API v1 to v2 consumer keys. The migration script updated the OAuth flow, but I forgot one thing: existing user connections were still using v1 tokens. The v1 key was deleted. So every Twitter post after March 11 silently failed with "Max retries exceeded."

How many posts failed? 22 posts across 4 users in 3 days. Including my own @BlogBurstAI account. The auto-pilot dutifully generated content, scheduled it, attempted to publish, failed, and moved on. No notification to me or the users.

The fix: Three code changes: (1) save consumer_key_version in OAuth credentials, (2) add a version property to TwitterOAuthService, (3) reconnect affected accounts with v2. Verified with a test tweet that went through instantly.

Silent Failure #2: The AI Brain Table Never Existed

This one hurt. The marketing_memories table — the entire foundation of the "AI that learns and improves" feature — was never created in the database. The model was defined. The service code was written. The learning processor was scheduled. But the table itself? Never created.

For 58 days, the system collected 4,056 learning events (post performance, engagement data, conversation insights). Every day, the learning processor tried to analyze them and save insights. Every day, it silently failed because the table didn't exist. Every day, it marked the events as "processed" anyway.

The agent was blind. It had the data, had the analysis code, had the Gemini integration — but nowhere to store what it learned. Every insight was lost.

The fix: One line: MarketingMemory.__table__.create(engine, checkfirst=True). Then ran the learning processor manually. Gemini analyzed the accumulated data and produced insights like "audience_insight content outperforms trends by 3.2x" and "Discord broadcasting completely fails." The AI finally has a working brain.

Silent Failure #3: The AI Was Replying "The Feeling Is Mutual!" Six Times

Checked the engagement logs. The AI had replied to one Twitter user (@HHH520527) six times in one day:

• "That means a lot. It's connections like this that we live for."
• "The feeling is mutual! We're so ready to ignite those possibilities."
• "It truly is. So glad we share that same wonderful feeling."
• "That's a beautiful point. This very exchange is a perfect example."
• "The feeling is mutual! So excited to see what you create."
• "It's happening! The feeling is mutual."

Six messages of pure nothing. Zero value. Maximum bot energy. No wonder nobody was following us.

The fix: Rewrote the reply prompt entirely. Old instruction: "Be warm and genuine." New instruction: "Lead with VALUE. Your first sentence must contain useful information, not pleasantries. NEVER start with 'That's so true' or 'The feeling is mutual.'" Also added per-author dedup (max 1 reply per person) and expanded the bot-phrase filter.

Silent Failure #4: Content Ignored Real Data

The auto-pilot was generating generic marketing hot takes: "Why are so many AI-driven content strategies failing despite massive investment?" This could have been written by anyone about anything. No real data, no personal story, no build-in-public authenticity.

Meanwhile, I had a goldmine of real metrics sitting in the database: 67 users, $15 MRR, 68 posts published this week, 22 failures. But the content generation prompt prioritized trending topics over product data.

The fix: Built a _get_build_in_public_data() function that pulls live metrics from the database and injects them into the content prompt as highest-priority context. The very next generated tweet:

"58 days in. 67 users. $15 MRR. It is definitely not a rocket ship yet, but the data is finally getting interesting. This week, 68 posts succeeded while 22 failed. That failure rate is where the AI actually learns."

That's a real tweet. With real numbers. From a real founder. That's what people follow.

Silent Failure #5: Posting to Wrong Platforms

User settings said "Twitter + Bluesky only." The auto-pilot was posting to Twitter, Bluesky, TikTok, Telegram, and Discord. Found the bug: line 345 had a comment that said // Always use all connected publishable platforms (ignore stale auto_pilot_platforms list). Someone (me) deliberately wrote code to ignore user preferences. Fixed.

Silent Failure #6: Thompson Sampling Was Learning, Nobody Was Listening

The Thompson Sampling algorithm — the multi-armed bandit that optimizes content strategy — was actually working. On Bluesky, it learned that casual tone has a 75% win rate vs technical tone at 25%. Bold claims as hooks work 62% of the time. Evening posts outperform morning.

But these learnings were being fed to the content generator alongside 1,034 "engagement_log" memories that had confidence=1.0 and crowded out the actual strategic insights. The top 20 memories by confidence were all useless reply logs.

The fix: Filtered the memory query to only load actionable types: winning_angle, failed_approach, audience_insight, platform_insight. Now the AI sees "casual tone works, trends don't" instead of a wall of reply logs.

Day 5 Numbers (March 14, 2026)

Users & Growth

• Total registered users: 67 (+0 today, +0 yesterday)
• Auto-pilot users: 7 (3 can actually post, 4 have broken Twitter connections)
• Paying customers: 0
• Website traffic: ~5 real visitors/day (rest is Googlebot)
• Bluesky followers: 11
• Twitter followers: single digits

Bugs Fixed

• Code changes: 4 files modified (auto_pilot.py, social.py, twitter_service.py, assistant_service.py)
• Database fixes: 1 table created, 3 records updated
• Emails sent: 3 (reconnect notifications to affected users)
• Deploys: 6+ (API + Celery worker restarts)
• Debugging time: ~4 hours with Claude Code

Competitor Research

• Direct competitors found: NoimosAI ($99-499/mo), Evatar.ai (video-focused)
• BlogBurst's niche: Only true autopilot + learning loop at less than $50/mo for indie hackers
• Key insight: Hypefury grew to $18K MRR through 15 months of daily Twitter activity, not product features

What I Learned

Silent failures are the startup killer. Not bugs that crash your app — those are easy to find. The deadly ones are features that look like they're working but aren't. My Twitter was "posting" (the scheduler ran, the logs looked normal). My AI was "learning" (events were collected, processor ran daily). My engagement system was "replying" (messages were sent). All technically true. All practically useless.

The lesson: monitor outcomes, not processes. "Did the tweet get posted?" is the wrong question. "Did the tweet get seen?" is better. "Did the tweet bring a follower?" is the right question. I was monitoring step 1 while steps 2-5 were broken.

67 users. 0 paying. 11 Bluesky followers. Day 58. The product works now — for real this time. The question is whether the internet notices.

Tomorrow's Focus

• Monitor Twitter engagement cycle — first time running with v2 key
• Track follower growth on Twitter and Bluesky over the next week
• Write the cold-start pain point article for IndieHackers and Reddit
• Continue optimizing SEO and GEO (AI search engine optimization)

Day 5. Six silent failures found and fixed. The AI agent finally has a working brain, a working voice, and a working pair of hands. Now it needs an audience. The hardest part of building a product isn't building it — it's getting anyone to notice it exists.