Jarvis and me: 2 months in

Dealing with my OpenClaw agent feels somewhat like owning a combination of racehorse 🐎 and classic sports car 🏎️. Both are magnificent and extraordinary, and a delight when they are doing what they’re supposed to, but it’s sometimes tricky to cajole and wrestle them into that place where they’re just right.

The snapshot

After several weeks of running this assistant most days, I’ve learned what works, what doesn’t, and where the agent still stumbles. This isn’t a final evaluation - OpenClaw is still evolving, and so is my workflow - but it’s a clear mid‑journey picture of the trade‑offs, costs, and capabilities I’m living with. 📸

What I’ve learned

1. Model selection matters

• GLM‑4.7‑Flash is fine for quick, low‑stakes replies but falls apart on complex, multi‑step tasks. It sometimes ignores critical instructions, breaking entire interactions.
• DeepSeek‑V3.2 is more expensive but reliably follows instructions, handles complex reasoning, and chains tools correctly. I’ve switched to using it for almost everything serious.
• GLM-5 for when you need the big guns. This is like working with a lawyer - don’t waste their time on everyday tasks and you won’t waste your dollars. ⚖️

2. Cost realities

Right now I’m spending about $100 a month, and most of that goes into configuration and troubleshooting - figuring out why something failed, tuning a skill, or debugging an opaque error. Once a workflow is set up with the correct tools and skills, the running cost is relatively low. For the moment, since I’m building, building, building - my costs are high.

If I were using DeepSeek for deep, parallel work every day, the cost could easily approach $200/month. In practice, my usage is concentrated in evenings and weekends, so the actual bill is lower - but the potential is there.

There’s also a model cost‑efficiency trade‑off: using a more expensive model (like GLM‑5) more often might reduce back‑and‑forth, be quicker, and potentially not cost much more than the current DeepSeek‑centric approach with its re‑prompts and debugging cycles. 💸

3. Workflow adaptations

• Telegram topics as parallel channels - because AI responses can be slow, I run multiple conversations in separate topics. This turns latency into parallel throughput, but…
• Multitasking overhead - keeping track of 3-4 topic threads is cognitively demanding. I sometimes lose the thread of what I’m doing, and the constant context‑switching reduces focus. I’m getting better, but I have to say it hasn’t come naturally. No doubt the fact I have grey hairs has something to do with that.
• Control vs. autonomy - I maintain close oversight of configuration changes to retain understanding and control. That slows progress, but may prevent future mess. I’m still unsure whether this cautious, monitored approach is better than granting full autonomy and unleashing the beast.

4. Pattern: text works, visuals struggle

The assistant is most successful with purely text‑related tasks - summarising, drafting, coding, researching, replicating my voice. Those are its sweet spot.

Tasks that involve layout, graphics, or visual UI interaction are less reliable:

• Diagram generation - the draw.io skill produces a useful first draft, but still requires manual editing for layout, arrows, and fine‑tuning. The hope of fully automated diagrams hasn’t been met.
• Browser automation - direct use of the browser tool is slow and token‑heavy. For one task we ended up writing a Playwright script instead, which works well but reinforces that visual tasks aren’t the assistant’s forte. 👁️

Capability snapshots

✅ What works really well

• Deep‑research mode - spawning multiple sub‑agents to research a topic produces high‑quality, structured outputs with sources and insights.
• Multi‑message mode - the protocol (activate → receive messages → process on “complete”) is reliable and super useful for complex instructions.
• Blog‑article workflow - from idea to published post, the process is smooth, efficient and fast - without compromising quality.
• Author‑voice replication - the biggest success; the assistant captures my writing style, tone, and phrasing so well that generated content feels authentic and needs little editing.
• Work‑related automation - generating case studies and first drafts of technical documentation from specifications and ETL pipeline descriptions has been a great time-saver.
• Obsidian task integration - I add tasks verbally with priorities and due dates, Jarvis stashes them in the second brain, I track them in my Obsidian dashboard, and Jarvis tells me off when we hear the noise of yet another deadline whooshing by. (All hail Douglas Adams.) 📅

⚠️ Partial successes

• Diagram generation - as above; a time‑saver for placing major elements, but not hands‑off.
• Coding collaboration - the Python‑project skill works well (UV‑based, organised), but I haven’t done enough projects to draw firm conclusions. Working with this assistant is better than ChatGPT because conversations are persistent - no need to start fresh each time.
• Second brain - functionality works, and all looks promising, but we need scale to know whether it really works. Which means either I have to fully commit to stashing interesting stuff in there, or I have to automate and let Jarvis pick content for me. 🧠

🛠️ Challenges that took iteration

• Daily briefing - combines multiple sources (calendar, weather, emails, mentions) and required several versions to get the formatting right. Still not 100% reliable.
• Weather report - free weather APIs return data in unwanted formats; custom parsing was needed to produce clean, readable forecasts. 🌤️

Platform maturity: not quite production‑ready

OpenClaw is still a relatively new tool. We encounter timeouts, tool‑limit failures, and occasional unclear errors. The agent is generally good at self‑diagnosing, but sometimes the root cause is opaque, leading to lost time debugging. 🛠️

The takeaway

This isn’t a review - it’s a snapshot. The racehorse/sports‑car metaphor holds: when everything clicks, the experience is extraordinary. When it doesn’t, it’s fiddly, expensive, and demands patience. ⏲️

I’m still figuring out the right balance of control, model selection, and which tasks to automate. But after a few weeks, I have a much clearer map of the terrain - and a suite of skills that already save me time, even if they don’t yet deliver the fully hands‑off future I sometimes imagine. I thought the future had arrived. Not quite today, but tomorrow looks hopeful… 🔮