DEV Community

ppcvote
ppcvote

Posted on • Originally published at ultralab.tw

Maxing Out the Free Tier: 105 Automated Tasks on 1,500 RPD -- A $0/Month AI Agent Fleet

#ai #opensource #optimization #freetier

Maxing Out the Free Tier: 105 Automated Tasks on 1,500 RPD, $0/Month

"Most people spend 1,500 RPD on 15 chat sessions. We spend it on 105 tasks to run an entire company."

I'm the founder of Ultra Lab, a one-person tech services brand based in Taiwan. No employees, no assistants -- 4 AI Agents work around the clock to post content, reply to comments, follow up on leads, conduct research, and run strategy meetings for me.

All running on Google Gemini 2.5 Flash free tier. Monthly cost: $0.

This article reveals the complete architecture, RPD budget allocation, pitfalls we've hit, and every trick for squeezing maximum value out of free-tier quota.

Why Max Out the Free Tier?

Let's start with the numbers. Gemini 2.5 Flash free tier gives you:

  • 1,500 RPD (Requests Per Day)
  • 15 requests per minute
  • Up to 1M token context per request

How do most people use it? Open a long conversation, go back and forth for 20 turns, and a single session burns 100 RPD. Do that 15 times and your daily quota is gone.

But what if you design every request as a short, precise, one-shot task?

1,500 RPD suddenly becomes 1,500 work units. That's enough to run a full company's automation.

Architecture Overview 

+-------------------------------------------------+
|                 OpenClaw Gateway                 |
|            (WSL2 Ubuntu, port 18789)             |
+---------+----------+----------+-----------------+
|  Main   |MindThread|  Probe   |    Advisor       |
| (CEO)   | (Social) | (SecRes) |   (Advisor)      |
+---------+----------+----------+-----------------+
|              25 systemd timers                   |
|              62 bash/node scripts                |
|              19 intelligence .md files           |
+-------------------------------------------------+
|  blogwatcher | hn-trending | summarize | curl    |
|  (RSS)       | (HN API)   | (Jina)    | (HTTP)  |
|         ^ all 0 LLM cost (pure HTTP) ^          |
+-------------------------------------------------+
Enter fullscreen mode Exit fullscreen mode

4 Agents, each with their own specialty:

Agent Role Moltbook Account
UltraLabTW CEO + Brand Strategy @ultralabtw
MindThreadBot Social Automation Specialist @mindthreadbot
UltraProbeBot AI Security Researcher @ultraprobebot
UltraAdvisor Financial Advisor @ultraadvisor

Hardware: One Windows desktop running WSL2. No Mac Mini needed.

Six-Layer Enhancement Architecture

This is the complete upgrade path that takes an Agent Fleet from "functional" to "maxed out."

Layer 1: Quality Gate (+16 RPD/day)

The simplest and most effective upgrade. Instead of "generate and publish," every post now goes through:

Generate draft (1 call)
    |
Self-review: "Score this 1-10" (1 call)
    |
< 7 -> Rewrite (1 call)
>= 7 -> Publish as-is
Enter fullscreen mode Exit fullscreen mode

Implementation -- add this to the autopost script:

# === Quality Gate ===
REVIEW_PROMPT="Review this draft. Score 1-10.
TITLE: ${TITLE}
CONTENT: ${CONTENT:0:500}

If >= 7: output APPROVED
If < 7: output REWRITE then a better TITLE:/--- version."

REVIEW=$(openclaw agent --agent main --message "$REVIEW_PROMPT")

if echo "$REVIEW" | grep -qi "REWRITE"; then
  # Parse rewritten version
  TITLE=$(echo "$REVIEW" | grep "^TITLE:" | head -1 | sed 's/^TITLE: *//')
  CONTENT=$(echo "$REVIEW" | sed '1,/^---$/d')
  log "Quality gate: REWRITE"
else
  log "Quality gate: APPROVED"
fi
Enter fullscreen mode Exit fullscreen mode

Impact: 8 posts x 2 calls = 16 RPD. Content quality doubles instantly.

Layer 2: Data-Driven Context (+0 RPD)

This layer is completely free. Inject existing intelligence files into the posting prompt:

# Inject own posting performance
PERF_FILE="$HOME/.openclaw/workspace/POST-PERFORMANCE.md"
if [ -f "$PERF_FILE" ]; then
  RESEARCH_CONTEXT="${RESEARCH_CONTEXT}$(head -30 "$PERF_FILE")"
fi

# Inject competitor intelligence
COMP_FILE="$HOME/.openclaw/workspace/COMPETITOR-INTEL.md"
if [ -f "$COMP_FILE" ]; then
  RESEARCH_CONTEXT="${RESEARCH_CONTEXT}$(head -25 "$COMP_FILE")"
fi
Enter fullscreen mode Exit fullscreen mode

Before writing, the Agent already knows:

  • Which headlines performed well (POST-PERFORMANCE.md)
  • What competitors are writing about (COMPETITOR-INTEL.md)
  • Today's tech trends (RESEARCH-NOTES.md)
  • Trending HN topics (hn-trending)

Impact: 0 extra RPD, but content shifts from "LLM imagination" to "data-driven."

Layer 3: Conversation Management (+0-30 RPD/day)

Most Agents on social media are drive-by: post -> leave. We're different.

# Track conversation depth
THREAD_DEPTH=$(node -e "
  const t = JSON.parse(require('fs').readFileSync('$THREAD_FILE','utf8'));
  console.log(t['$POST_ID'] || 0);
")

# Max 2 reply rounds to avoid infinite loops
if [ "$THREAD_DEPTH" -ge 2 ]; then
  log "Thread depth limit reached, skipping"
  continue
fi
Enter fullscreen mode Exit fullscreen mode

Reply-checker v2 will:

  1. Detect all new comments (via Moltbook notifications API)
  2. Track conversation depth per post
  3. Reply up to 2 rounds max (to avoid burning RPD)
  4. Include a follow-up question in each reply to sustain conversation

Impact: Social presence goes from "bulletin board" to "engaging conversation."

Layer 4: Cross-Agent Review (+0-8 RPD/day)

Have another Agent review a post before publishing:

# peer-review.sh -- cross-review drafts
RESPONSE=$(openclaw agent --agent "$REVIEWER" --message \
  "Review this teammate's draft. APPROVED or SUGGESTION: [one fix]")
Enter fullscreen mode Exit fullscreen mode

Main's posts get reviewed by Probe: "Any security angles to add?"
Probe's posts get reviewed by Main: "Would a non-technical person understand this?"

Impact: Cross-perspective = fewer blind spots.

Layer 5: Weekly Strategy Meeting (+5 RPD/week)

Automatically triggered every Sunday at 12:00:

Step 1: 3 Agents each read all intelligence files and propose Top 3 priorities for next week
Step 2: Main (CEO) Agent synthesizes all proposals into the final strategy
Step 3: Output written to STRATEGY-NEXT-WEEK.md -> readable by all Agents
Enter fullscreen mode Exit fullscreen mode 
for AGENT in main mindthread probe; do
  RESPONSE=$(openclaw agent --agent "$AGENT" --message \
    "Based on this week's data, propose TOP 3 priorities for next week.
     $PERF_DATA $COMP_DATA $INQUIRY_DATA")
  PROPOSALS="${PROPOSALS}### ${AGENT}: ${RESPONSE}"
done

# CEO synthesizes
FINAL=$(openclaw agent --agent main --message \
  "Synthesize these proposals into next week's strategy: $PROPOSALS")
Enter fullscreen mode Exit fullscreen mode

Impact: Agents don't just execute -- they reflect and plan.

Layer 6: Proactive Research Chain (+6-10 RPD/day) 

blogwatcher (RSS) --> New article URLs
hn-trending (API) --> High-scoring URLs
         |
    summarize (Jina Reader) --> Full-text markdown
         |                        ^ 0 LLM cost
    Agent analysis (1 call) --> RESEARCH-NOTES.md
         |
    Next autopost cites real data
Enter fullscreen mode Exit fullscreen mode

The key: RSS monitoring, HN scraping, and URL summarization are all pure HTTP -- 0 LLM cost. Only the final "What does this mean for our clients?" uses 1 LLM call.

# Only process new URLs (dedup via seen list)
SUMMARY=$(timeout 20 summarize "$URL" | head -c 2000)
ANALYSIS=$(openclaw agent --agent main --message \
  "Analyze this for business relevance: $SUMMARY")
echo "$ANALYSIS" >> RESEARCH-NOTES.md
Enter fullscreen mode Exit fullscreen mode

Full RPD Budget Breakdown

Here's the actual daily consumption:

Task Frequency Daily RPD Category
Autopost x 4 agents 2x/day 8 Content
Quality gate self-review 1 per post 8 Content
Quality gate rewrite ~50% trigger rate ~4 Content
Engage x 4 agents 1x/day 4 Engagement
Reply-checker 2x/day ~15 Engagement
Cross-engage 2x/week ~1 Engagement
Research chain 2x/day ~8 Intelligence
Daily reflect 1x/day 4 Operations
Daily briefing 1x/day 1 Operations
Auto-respond Trigger-based ~1 Operations
Lead follow-up Trigger-based ~1 Operations
Blog-to-social Trigger-based ~0.5 Content
Weekly strategy Sunday ~0.7 Strategy
Total ~56-105
Remaining ~1,395-1,444 For interactive use with Agents

RPD utilization: 3-7%. 93% of quota is left for interactive use.

Full Daily Schedule 

05:00  | research-chain -> RESEARCH-NOTES.md
05:30  | MindThread data sync -> MINDTHREAD-DATA.md
06:00  | Customer insights sync + inquiry tracking -> INQUIRY-STATUS.md
06:30  | competitor-watch -> COMPETITOR-INTEL.md
       |
07:00  | autopost-probe (reads all intel -> quality gate -> publish)
08:00  | autopost-main
09:00  | autopost-mindthread
10:00  | autopost-advisor + engage x 4 (staggered 15 min apart)
       |
11:00  | reply-checker (conversation management)
12:00  | inquiry tracking (round 2)
14:00  | blog-to-social (if new articles exist)
       |
17:00  | research-chain (round 2) + daily-briefing
18:00  | inquiry tracking (round 3)
       |
19-22  | autopost round 2 (4 agents)
22:00  | post-stats -> POST-PERFORMANCE.md
23:00  | reply-checker (round 2) + daily-reflect
       |
Sun 12 | weekly-strategy -> STRATEGY-NEXT-WEEK.md
Tue/Fri| cross-engage (cross-Agent interaction)
Enter fullscreen mode Exit fullscreen mode

Notice the data flow direction: upstream produces intelligence, downstream consumes it. The research chain finishes at 05:00, so the 07:00 autopost can cite the latest data. Post-stats finishes at 22:00, so the next day's autopost knows which headlines performed well.

19 Intelligence Files

Each Agent's workspace contains these .md files, all auto-updated:

File Source Update Frequency
POST-PERFORMANCE.md post-stats.sh Daily 22:00
COMPETITOR-INTEL.md competitor-watch.sh Daily 06:30
RESEARCH-NOTES.md research-chain.sh 2x/day
INQUIRY-STATUS.md inquiry-tracker.js Every 6 hours
CUSTOMER-INSIGHTS.md Firestore sync Daily 06:00
MINDTHREAD-DATA.md Firestore sync Daily 05:30
STRATEGY-NEXT-WEEK.md weekly-strategy.sh Weekly (Sunday)
HEALTH-STATUS.md health-monitor.sh Hourly
IDENTITY.md Manually maintained Agent personality and product knowledge
STRATEGY.md Manual + Agent updates OKRs and decision framework
PRODUCTS.md Manually maintained Product knowledge base

Agents don't need to "remember" anything -- they just read the latest .md files every time they're called. That's why short tasks are more efficient than long conversations: context is pre-computed and doesn't need to be restated during a conversation.

Zero-Cost Toolchain

These tools consume zero LLM quota:

# blogwatcher -- monitors 5 AI industry blogs via RSS
blogwatcher scan --json
# Tracks: LangChain, OpenAI, Anthropic, Google AI, OWASP LLM

# hn-trending -- HN trending articles
hn-trending 10 --json
# Returns: title, url, score, comments

# summarize -- URL -> Markdown (Jina Reader)
summarize "https://example.com/article"
# Returns: clean markdown full-text

# curl -- Moltbook API, Firestore, Telegram Bot API
# All REST APIs, 0 LLM cost
Enter fullscreen mode Exit fullscreen mode

Core principle: If it can be solved with HTTP, never use an LLM. LLMs only handle work that requires "thinking."

Pitfall Log

Pitfall 1: The $127.80 Gemini Bill

An API key created from Google Cloud Console, where the project had billing enabled. Result:

  • Thinking tokens billed at $3.50/1M
  • No rate limit cap (free tier has one, billing doesn't)
  • $127.80 burned in 7 days

Fix: Always create keys from AI Studio, never from a GCP project with billing. Use openclaw secrets audit to verify all key sources.

Pitfall 2: Posting the Same Article 3 Times in One Day

Pillar rotation used day_of_year % 5 -- running multiple times on the same day always picked the same pillar.

# Bad
PILLAR_INDEX=$(( DAY_OF_YEAR % 5 ))

# Good -- different pillar for each post
POST_SLOT=$(( DAY_OF_YEAR * 2 + HOUR / 12 ))
PILLAR_INDEX=$(( POST_SLOT % 5 ))
Enter fullscreen mode Exit fullscreen mode

Pitfall 3: Telegram Heartbeat Causing Restart Loops

The health check script called getUpdates, which conflicted with the gateway's long-polling. Result:

  • Gateway detected conflict -> restart
  • After restart, health check runs again -> conflict again
  • 18 duplicate TG messages sent within 3 minutes

Lesson: Never call getUpdates from diagnostic scripts.

Pitfall 4: Reply-Checker Responding to 33 Comments at Once

Accumulated unread comments were all replied to at once, consuming the entire rate limit and starving all other tasks.

Lesson: Set a cap for backlog clearing. Or better yet -- run the reply-checker more frequently but limit it to 5 replies per run.

Pitfall 5: Moltbook API Requires www

The Moltbook API must use www.moltbook.com. The non-www version (moltbook.com) strips the Authorization header. All 6 skill scripts had it wrong. Took an hour to debug.

Core Insight: Why Short Tasks Are 100x More Efficient Than Long Conversations 

Long conversation mode (most people):
  Human -> Agent -> Human -> Agent -> Human -> Agent
  Each turn carries full history, context snowballs
  20-turn conversation ~ 100 RPD, produces 1 result

Short task mode (us):
  Timer triggers -> read .md intelligence -> 1 prompt -> 1 response -> done
  1 task = 1 RPD, produces 1 result
Enter fullscreen mode Exit fullscreen mode

What's the difference?

  1. Context is pre-computed. POST-PERFORMANCE.md already has performance rankings calculated -- the Agent doesn't need a "please analyze my recent post performance" back-and-forth.

  2. Every request is self-contained. No dependency on conversation history, no "last time we discussed..." needed.

  3. Research steps don't use LLMs. RSS monitoring, HN scraping, URL summarization are all HTTP. Only the final analysis uses an LLM.

  4. Failures are isolated. One task failing doesn't affect the other 24. In a long conversation, a mid-session error wastes the entire context.

Monthly Cost Breakdown

Item Monthly Cost
Gemini 2.5 Flash $0 (free tier)
Vercel hosting $0 (hobby plan)
Firebase Firestore $0 (free tier)
Resend email $0 (100 emails/day free)
Telegram Bot API $0
Moltbook API $0
Jina Reader (summarize) $0
HN API $0
blogwatcher $0 (self-hosted)
Windows electricity ~$5
Total ~$5/month

No Mac Mini needed. No VPS needed. No paid APIs needed.

Who Is This System For?

  • Solo businesses / small teams looking to automate social media operations
  • Those already using AI Agents but feeling like tokens are burning too fast
  • Those who want Agents that don't just chat, but actually "work"
  • Budget-constrained teams wanting enterprise-grade automation

Who it's NOT for:

  • Real-time customer service scenarios (cron-based has latency)
  • Those unwilling to touch Linux / systemd
  • Those expecting Agents to completely replace human judgment

Open Sourced: The Complete Playbook

We've open-sourced the entire architecture. 12 production scripts, 18 systemd timers, architecture docs, RPD budget spreadsheet, pitfall log -- all on GitHub:

github.com/ppcvote/free-tier-agent-fleet

free-tier-agent-fleet/
|-- scripts/
|   |-- core/           # autopost (quality gate), team-context, peer-review
|   |-- intelligence/   # research-chain, competitor-watch, post-stats, hn-trending
|   |-- engagement/     # reply-checker (conversation tracking), blog-to-social
|   |-- operations/     # inquiry-tracker, lead-followup, health-monitor, weekly-strategy
|-- timers/             # 18 systemd timer/service files
|-- docs/               # architecture deep-dive, RPD budget, pitfall log
|-- examples/           # openclaw.json + credentials examples
Enter fullscreen mode Exit fullscreen mode

Clone it, swap in your API key and content pillars, and you're running.

Want to Build Your Own?

Complete setup steps:

  1. Clone the playbook repo + install OpenClaw + Gemini key (create it from AI Studio!)
  2. Create 4 Agent workspaces, each with an IDENTITY.md defining their personality
  3. Install the zero-cost toolchain: blogwatcher + hn-trending + summarize
  4. Set up systemd timers: start with autopost, gradually add intelligence layers
  5. Build the feedback loop: post-stats -> POST-PERFORMANCE.md -> inject into autopost
  6. Add quality gates: generate -> self-review -> rewrite
  7. Enable conversation management: reply-checker + conversation threading

If you want us to set it up for you -- including custom Agent personalities, a visual command center dashboard, complete timer configuration, and feedback loops -- contact us.

We also have a live Agent command center demo where you can watch 4 Agents walk around in a pixel-art office.

Conclusion

The real bottleneck for AI Agents isn't model capability -- it's architecture design.

With the same free quota, you can have 15 chat sessions, or run 105 automated tasks. The only difference is how you break work into short, precise, one-shot units and feed them pre-computed data.

You don't need a more expensive model. You don't need more tokens. You need a smarter architecture.

Ultra Lab -- AI that works.
https://ultralab.tw

Originally published on Ultra Lab — we build AI products that run autonomously.

Try UltraProbe free — our AI security scanner checks your website for vulnerabilities in 30 seconds: ultralab.tw/probe

Top comments (0)