September 2026 Major Update: Smarter Agents, X Feed Coverage, CEO Brain Upgrades & 30+ Fixes

The Short Version

Over the past week we shipped 30+ improvements across the server, the free plugin, and the Pro plugin. The big themes:

• 1.Agents that learn from mistakes — every quality failure becomes a lesson the writer remembers
• 2.CEO Brain that actually knows your fleet — richer context, persistent preferences, proactive warnings
• 3.X Feed Follower — follow any Twitter reporter, approve scoops with one tap
• 4.Aggressive article recovery — no more 78-word garbage drafts
• 5.12+ quality gate fixes — stale generations, broken keywords, blank links, glued FAQs

Here's everything, organized by what it means for you.

---

1. Agents That Learn From Their Mistakes

This is the feature we're most excited about. Every AI agent now has a memory loop:

How it works

When an article gets blocked by a quality gate (truncated, too short, stale content, bad keywords), the failure reason is recorded as a "mistake lesson" in your account's brain. The next article written for that category automatically receives a prompt injection:

RECENT MISTAKES IN THIS CATEGORY — DO NOT REPEAT

- Produced truncated paragraphs — end every section with complete sentences

- Topic was stale — covered an event from August, not current

The same mistake seen again reinforces the lesson (effectiveness score rises toward 0.9) instead of flooding your database with duplicates.

Why this matters

Before this change, an agent that produced a truncated article would just... produce another truncated article on the next cycle. The quality gates caught it and drafted it, but nothing told the writer to stop doing that. Now the writer literally sees "don't do this" in its prompt.

---

2. CEO Brain That Actually Knows Your Fleet

The CEO Brain (your AI chief of staff) received a massive intelligence upgrade:

Persistent Preferences

Tell the CEO "always write 800 words for tech articles" or "never auto-publish entertainment drafts" — it saves this as a permanent preference that the writer applies to every future article. This persists across ALL conversations and sessions.

Proactive Intelligence

The CEO now sees your best and worst performing articles with scores, recent quality failures, and performance priors. It will proactively say things like:

"I noticed your gaming articles underperform — want me to try shorter formats?"

Instead of waiting for you to ask.

New Tools

• set_preference — save standing instructions the writer always follows
• manage_drafts — publish or delete drafts from chat
• update_agent — edit interval, word count, niches, publish window from chat
• bulk_control — pause or resume ALL agents at once
• cover_url — paste any article URL, the CEO fetches it and writes a grounded story

---

3. X Feed Follower — Never Miss a Scoop

Follow any Twitter handle. New posts are scouted every 15-30 minutes and either sent to you as Telegram approve-cards (✅ Cover / ⏭ Skip) or auto-covered.

What's new this cycle:

• Freshness gate: only tweets from the last 15-60 minutes (matching your poll interval) trigger cards — no more 11-hour-old "news"
• Priority execution: approved stories jump the queue and write immediately, not on the agent's next hourly cycle
• Self-dedup: overlapping variants of the same story are caught before dispatch
• Domain routing: videocardz.com → tech agent, cricbuzz.com → cricket agent — no more arbitrary agent assignment

Read the full X Feed Follower announcement.

---

4. Aggressive Article Recovery

The #1 cause of failed articles: the AI model returning severely short content (78 words on a 600-word target). This happened because minimax-m3 (the free pool model) sometimes collapses mid-generation.

What we fixed:

• Three-layer fallback: if the primary model produces short output → retry with deepseek → if still short → retry with gemini → only then 422
• Widened mid-band recovery: articles in the 5-50% range now trigger the fallback (was only <25%)
• Accept ANY improvement: a 200-word deepseek article beats a 78-word minimax one — the old code required 2× improvement
• 80/20 pool reweight: deepseek (reliable) gets 80% of pool articles, minimax 20%

---

5. Quality Gate Improvements

Twelve fixes to the quality pipeline:

---

6. Model Reliability

• 80/20 pool reweight: deepseek-v4-flash (reliable workhorse, $0.09/$0.18 per 1M) now writes 80% of pool articles. minimax-m3 dropped to 20%
• Topic format-nudge retry: when the topic model returns empty/malformed, retry with deepseek instead of the same model
• Aggressive short-output recovery: three-layer fallback (primary → deepseek → gemini) before 422ing
• Model-metadata salvage: "google/... Generated Content" titles get a deepseek retry before rejection
• Guaranteed topic terminal fallback: duplicate exhaustion no longer kills the cycle — the guaranteed bank always produces a topic

---

Getting Started

• 1.Update the free plugin to v2.5.343+ (download)
• 2.Update the Pro plugin to v2.4.77+ (from your dashboard)
• 3.Try the new CEO commands: "set preference: always 800 words for tech", "show my worst articles", "cover "
• 4.Follow your first X feed: /feeds add @FabrizioRomano football ask 30

Questions? The CEO Brain knows everything about your setup — just ask.