OpenClaw Six-Month Retrospective: $0, 4 AI Agents, 35 Scheduled Jobs
Table of Contents
- Every number on the table
- System scale
- What surprised me
- 1. Output beyond what I imagined
- 2. Quality better than expected
- 3. Cross-agent collaboration really happens
- What disappointed me
- 1. Deep content still has to come from me
- 2. Some platforms are not suited to automation
- 3. Maintenance cost is not zero
- RPD budget allocation (real data)
- Error rates
- Time saved
- Where the limits of a one-person company are
- What an AI team cannot do
- The real limit of a one-person company
- If I did it again
- For anyone who wants to start
- Week one: one agent, one task
- Week two: add a quality gate
- Week three: add intelligence sources
- Week four: consider a second agent
- After that: expand slowly
- The full series index
- Finally
Disclosure: Ultra Lab, MindThread and Ultra Advisor are products of the same group, Ultra Creation (傲創實業).
Update, 27 September 2026. The numbers and system scale in this post are as they stood when it was written in April 2026. OpenClaw's agent functions were retired in June 2026 and the fleet moved to Hermes Agent, with model calls now routed through OpenRouter; the "$0 a month on the Gemini free tier" figures describe how it worked at the time. We later ran an adversarial audit on this fleet: of the 14 chapters of Agentic Design Patterns we claimed to have implemented, not one came out completely clean.
Every number on the table
In the first five posts of this series, I covered why, how, and what went wrong.
In this last one, I want to do something many people won't: publish six months of real data, all of it.
The good, the bad, what surprised me, what disappointed me. All of it.
System scale
First, the size of the system as of 2026-04-14:
| Metric | Number |
|---|---|
| Agents | 4 (Main, MindThread, Probe, Advisor) |
| Systemd Timers | 35+ |
| Daily automated tasks | 100+ |
| Daily automated posts | 16+ |
| Daily automated replies | 20-40 |
| Daily intelligence collection | 2 runs (RSS + HN) |
| Daily competitor analysis | 1 run |
| Daily health checks | 24 (hourly) |
| Weekly strategy meeting | 1 (the agents hold it themselves) |
| Monthly cost | $0 (Gemini free tier) |
| RPD usage | ~5.3% (79/1500) |
What surprised me
1. Output beyond what I imagined
Six months ago, working alone, I could write at most 1-2 articles a day, and after that I had no energy left for anything else.
Now the system produces 16+ posts a day automatically, from 4 different angles. Six-month totals:
- Moltbook posts: 2,800+
- Blog articles: 125+ (bilingual, Chinese and English)
- Automated replies: 3,000+
- Intelligence reports: 360+
What if I did all of this myself? At 30 minutes a post, the posting alone would take 8 hours a day. The maintenance time I actually spend on OpenClaw: 15 minutes a day on average.
2. Quality better than expected
After adding quality gates, the agents' article quality settled at 7-8 out of 10.
What surprised me more is that several articles got better engagement than ones I wrote myself. Possible reasons:
- Agents are more consistent: quality does not drop because I am in a bad mood that day
- Agents have intelligence: they read HN and RSS every day and know better than I do which topics are hot right now
- Agents do not procrastinate: when the schedule fires, they write. No "I don't feel like it today, tomorrow then"
3. Cross-agent collaboration really happens
After the Content Relay mechanism went live, I started seeing spontaneous "relays" between agents.
The classic example: Probe published a technical analysis of the OWASP Agentic Top 10. In the next round, Main read the relay and decided on its own to write a follow-up from the angle of "why enterprises should care". Then MindThread made a summary graphic card.
Three posts, three angles, fully automatic. I only saw them afterwards.
In that moment I thought: this is not just automation anymore, this is collaboration.
What disappointed me
1. Deep content still has to come from me
Agents are very good at: trend analysis, news commentary, organising knowledge, formatted content.
Agents do poorly at: deep reflective pieces like this series. Content that needs personal experience, emotion and real stories comes out of AI missing that particular flavour.
My current split:
Agent does: everyday content (80%)
→ trend analysis, technical summaries, knowledge sharing, community interaction
I do: deep content (20%)
→ personal experience, strategic reflection, product decisions, this series
The 80/20 rule. The agents carry the daily volume, and I focus on what only I can write.
2. Some platforms are not suited to automation
We paused automated posting on Bluesky in March: the API has a facet URI bug, with a 100% failure rate.
Some platforms have very strict anti-spam mechanisms. However well an agent's behaviour imitates a human, it still gets caught.
What we learned: not every platform is suited to automation. Picking the right battlefield matters more than technical ability.
3. Maintenance cost is not zero
The monthly fee is $0, but the time cost is not.
- Expired tokens need handling (I fixed this bug just last week)
- Banned accounts need handling
- Platform API changes need to be followed
- Quality occasionally slips and the prompt needs tuning
It averages 15 minutes a day, but now and then a problem blows up that takes hours (like that $127.80 bill incident).
Automation is not "set it and forget it". It is "going from the person who does the work to the person who manages the system". The management cost is lower, but it is not zero.
RPD budget allocation (real data)
The Gemini free tier (as of 2026-04-14) allows 1,500 RPD a day. How much we actually used:
| Category | RPD/day | Share |
|---|---|---|
| Content generation (4 agents × ~3 calls × 2 runs/day) | 24 | 1.6% |
| Automated replies | 34 | 2.3% |
| Intelligence research | 15 | 1.0% |
| Operations (health checks, customer follow-up) | 5 | 0.3% |
| Weekly strategy | 1 | 0.07% |
| Total | 79 | 5.3% |
94.7% headroom remaining.
This means:
- The system can scale 10x without exceeding the free tier
- Even if a bug causes mass retries one day, it will not blow up
- The Gateway's 1,500 RPD hard cap is the last line of defence
Many people ask me "is the free tier enough?" Not only is it enough, we have not even used 10% of it.
Error rates
Error tracking over six months:
| Type | Count | Impact |
|---|---|---|
| Token expiry | 12 times | Account idle for 1-4 days |
| Account bans | 8 times | Had to wait for the ban to lift |
| Content below the quality bar | ~30% of first generations | Stopped by the quality gate and rewritten |
| System downtime | 3 times | Recovered within 2 hours on average |
| API changes | 2 times | Required manual code fixes |
| Serious incident | 1 time | $127.80 bill |
System availability: ~99% (excluding downtime and token issues)
The most common problem is token expiry. We have fixed the blind spot in automatic refresh (covered in the second post of this series), so it should drop sharply from here.
Time saved
The most important number: how much time does this system actually save me?
| Task | Time before | Time now |
|---|---|---|
| Daily content production | 2-3 hours | 5 minutes (reading reports) |
| Community interaction | 1 hour | 5 minutes (checking anomalies) |
| Market research | 30-60 minutes | 0 (automatic) |
| System monitoring | Not done (I found out only when something broke) | 0 (automatic alerts) |
| Competitor tracking | Only when it occurred to me | 0 (automatic, daily) |
| Total | 4-5 hours/day | 15 minutes/day |
4+ hours saved every day. What do I use that time for?
- Writing UltraProbe code and sending PRs to Cisco and Microsoft
- 1:1 voice chats with Discord community members
- Thinking about product strategy
- Writing this series (deep content only a human can write)
The point is not "how much time was saved", it is "where the time goes". The AI team took me from "doing everything and doing none of it well" to "focusing on the highest-value 20%".
Where the limits of a one-person company are
After all these upsides, let's end with the limits.
What an AI team cannot do
Build real relationships: the Discord community grew from 0 to 268 people through my weekly voice chats and 1:1 conversations, not through the agents' automated replies
Make product decisions: whether to add this feature, whether to cut that product line, how to set prices. Only a human can make these calls
Handle the unexpected: when an application is rejected, when a bill spikes, when a partner suddenly changes their mind. AI will not handle your emotions or make the hard decisions for you
Create something truly original: agents can organise, analyse and extend, but they cannot create a new concept from scratch. OpenClaw itself is an architecture a human came up with
The real limit of a one-person company
It is not how much you can get done. The AI team has already solved the problem of "not enough hands".
The real limit is attention.
Even if AI does 80% of the work for you, if the remaining 20% is spread across 6 product lines, each one only gets 3.3% of your attention.
An AI team lets a one-person company do what used to take 5-10 people. But it will not turn you into a superhuman. You still have one brain, one supply of attention, and 24 hours in a day.
My solution: unequal allocation. 60% of my attention goes to UltraProbe (top priority), 30% to MindThread (the revenue hope), and 10% is split across the rest. The AI team handles the things that "need doing but do not need my full focus".
If I did it again
If I could start over, what would I change?
Start earlier: I hesitated too long, worried that AI quality would not be good enough. The truth is that quality can be iterated on, but time does not come back
Build monitoring before automation: several disasters happened because "it broke and I didn't know". My rule now: when a script is written, write its health check
Do not chase perfection: publish at 7, do not grind it to 9. Consistency matters more than occasional perfection
Add one agent at a time: I launched all 4 at once, and when debugging I had no idea which one the problem was in. I should have stabilised one before adding the next
Record everything: I only started the war stories document later. If I had recorded every problem and its fix from the start, I would have avoided a lot of repeated mistakes
For anyone who wants to start
If you are also one person carrying too much and want to try an AI team, here is my advice:
Week one: one agent, one task
Do not take on too much at once. Pick the most painful repetitive task (for example, publishing one LinkedIn post a day) and automate it with one agent.
Week two: add a quality gate
After the agent has run for a week, you will know where its problems are. Add self-review and tune the prompt.
Week three: add intelligence sources
Stop the agent generating from thin air, and have it produce output based on real data.
Week four: consider a second agent
Once the first one is stable, add another. Do not rush.
After that: expand slowly
Add monitoring, add collaboration mechanisms, add strategic planning. Make sure each step is stable before moving on to the next.
The full series index
- Why I needed an AI team: the journey + a decision framework
- The day my first agent went live: 6 pitfalls + Safety Patterns
- A zero-cost intelligence system: RSS + HN + Jina pipeline
- Multi-agent collaboration architecture: a 6-layer collaboration design
- Quality gates in practice: Self-Review + Pillar Rotation
- Six-month retrospective (this post): real data + reflections
All source code is on GitHub: free-tier-agent-fleet
Finally
Six months ago, I was a one-person company founder about to be crushed by my own ambition.
Today I am still one person. But I have 4 AI teammates that never sleep, never take leave, and cost $0 a month.
They are not perfect. They make mistakes, produce garbage, and let tokens expire. But they also took me from "doing everything and doing none of it well" to being able to spend my time on what truly matters.
A one-person company + an AI team is not the future. It is now.
And the barrier to entry is much lower than you think.
If this series helped you, come talk to me on Discord, or fork the GitHub repo and start your own AI team.