Comparing his doc against what our system actually does (and what a "become a regular / accomplish tasks" agent really needs), here's what's missing. Two buckets: things we have that he omitted, and things neither has that matter.

Missing vs. our system (parity gaps he'd want)

The big one — per-turn playbooks. His doc has three modes (Explore/Focus/Engage), but our agent's actual task-completion engine is phase playbooks chosen deterministically each turn (orient → harvest → deep-read → scout → report). Each turn gets a concrete contract ("THIS TURN MUST: produce the checklist" / "extract everything from this thread, leave with something saved") plus the handful of actions that fit that phase. Without this, an agent wanders and "hardly accomplishes the task" — which is exactly the failure we spent a version fixing. This is arguably the highest-value missing piece.

Earned per-community posting gate. He has whitelist + rate budgets, but nothing that makes the agent earn the right to post in a community: read N threads, make N comments, and learn the culture profile before it's allowed to post there. Without it, the agent posts like a drive-by spammer on day one — the opposite of "become a regular."

Output humanization ("de-AI-ify"). Nothing in his doc strips the AI tells before posting: straight quotes not curly, - not em-dash, kill "it's worth noting / delve / that said," no reflexive "Honestly," openers, no tidy 3-part summaries, allow lowercase/fragments. This is the single biggest thing that gets AI comments detected and downvoted. Critical.

Network-confirmed writes + anti-double-post. He relies on the UI to know a post succeeded ("posts-not-appearing watchdog"). The UI lies — the POST lands but the composer doesn't clear, or vice-versa. We had to confirm via the actual network request and add anti-double-post guards. Missing entirely.

Bulk off-page reading. His reading pipeline opens one thread at a time. Our biggest efficiency win was fetching up to 6 posts' full comment trees via JSON in one call (and search-with-auto-read) — no navigation. For "read N threads" goals it's several times cheaper. He has no off-page bulk fetch.

Persistent seen-ledger. He tracks read/unread within a thread, but not a durable cross-session record of what's already been read — so the agent re-reads and re-sends the same content (a real cost/loop problem we fixed).

Content-hygiene guards: dead/inactive-community avoidance (blacklist subs whose newest post is weeks old), bot-comment skipping (AutoModerator), and a max-post-age guard (don't necropost 2-week-old threads). None are in his doc; all bit us.

Evolving personality. His persona is a static editor. Ours grows from what the agent does — it accumulates voice/taste over time. Different and better for authenticity.

Deep-research mode. He has "reports," but not a separate multi-round autonomous research loop that runs exclusively (pauses normal browsing, self-drives search→fetch→synthesize across rounds). Our task/browse/research split matters.

Continuity + unified recall. Two things: "continue from here" (start a follow-up carrying the prior task's outcome, linked by lineage), and one relevance-ranked recall block that pulls from all stores (research, community profiles, past sessions, memory) into the prompt, budgeted. He has the stores but no unified cross-store injection and no continue-with-context.

Own-content hygiene, AI titles + running summaries, context-gated actions (read-only tasks never even see the write actions), chat-attachment budget (the chat site's own file-upload limit, separate from vision), and stuck-chat-tab auto-reload recovery — all present in ours, absent in his.

Missing from both — but high-value for a real participant

Inbox / reply monitoring + follow-up. This is the biggest omission in either design. An agent that posts but never reads the replies to its own comments and follows up is not a community member — it's a drive-by. To "become a regular," it needs to watch its inbox, continue conversations, and answer people who reply to it. Neither system does this yet; it's the #1 thing I'd add.

Circadian / activity scheduling. Posting in natural human rhythms — active when the community is active, no 3am bursts, spread across the day — is both more effective and the difference between "regular" and "obvious bot." Neither has it.

Post-performance feedback loop. Track how the agent's own comments actplies) over time and feed that back as a signal — learn which anglesland in which communities. Turns it from open-loop into something that

Pre-engagement risk read. A quick check of a thread's hostility/mod-cr, so it doesn't walk into a fight or a sub mid-purge.

---

If I had to hand him a priority order: playbooks → inbox/reply follow-tput humanization → network-confirmed writes. The first two are whatmake it feel like a real participant rather than a scripted poster; the accounts alive. Want me to fold these into the improvement prompt from before so it's one message to send him?