From OpenClaw to Claude Code: Our Journey in Agentic AI in 2026

Neto Pompeu - CEO Primo Studio —

From OpenClaw to Claude Code: Our Journey in Agentic AI in 2026

I'm writing this article in the first person because this is our experience at Primo Studio. An honest feedback on what we built, what worked, what broke, and where we stand today in agentic AI.

January 2026: Starting with OpenClaw

In January, we set up OpenClaw with Claude Opus 4.6. For those discovering it, OpenClaw was an open source harness allowing the use of Anthropic models with full tooling and an agentic architecture. For an agency like ours based in French Guiana, it was perfect: we could build custom agents without relying on an opaque American SaaS platform.

We used it intensively, to the point it became central to our workflow:

  • Code generation
  • Automation of administrative tasks
  • Client document analysis
  • Rapid prototyping

Our Contribution: the OpenClaw Memory Plugin

One major missing feature of OpenClaw back then was memory persistence between sessions. Every conversation started from scratch, the agent didn’t remember your preferences or the business context of your projects.

So, we developed a dedicated memory plugin: openclaw-memoria.

What it did:

  • Persistent project context storage
  • Layered memory (short-term, long-term, meta)
  • Smart retrieval based on relevance
  • Context sharing among agents

This marked the beginning of our serious exploration of agentic AI in production.

The Qwen 35B Experience on Mac mini 64GB

At the same time, we wanted to test sovereignty. No way were we going to bet everything on a cloud API that can vanish overnight (spoiler: this proved prophetic).

We deployed Qwen 3.5 35B on a Mac mini M4 Pro 64GB RAM. The figures were impressive:

  • Local inference at 15-20 tokens/second
  • Very acceptable output quality for simple tasks
  • Nearly zero marginal cost after hardware purchase
  • 100% local data, zero cloud sending

The real advantage: for repetitive workflows (email writing, summaries, basic translations), local Qwen was more than enough. No need to pay $15 per million tokens.

The WhatsApp Admin Assistant

We started setting up an administrative assistant via WhatsApp Business API, connected to our stack:

  • WhatsApp → webhook
  • Processing by Claude Opus via OpenClaw
  • Access to tools: calendar, Directus, billing, CRM
  • Automated client responses

The idea: allow a SME manager in French Guiana to manage their business via WhatsApp, with no complex dashboard. “Send an invoice to Dupont for €500” → done.

The Hard Blow: Anthropic Stops OpenClaw

Then in March/April 2026, Anthropic decided to stop supporting OpenClaw. For strategic reasons (likely consolidation around Claude Code), the API access dedicated to OpenClaw was cut.

This forced us to pivot. Quickly.

The GPT-5.4 Attempt: Good but Not Up to Par for Us

Naturally, we tested OpenAI’s GPT-5.4 as an alternative. The verdict after several weeks of real use:

Positives:

  • Good overall quality
  • Stable API
  • Reasonable pricing
  • Long context

Blocking points for us:

  • Very limited tooling: GPT struggles when chaining multiple tools precisely. Claude Opus chained 15-20 tool calls while maintaining context; GPT-5.4 gets lost after 5-6.
  • Excessive verbosity: responses are always long, full of disclaimers and preambles. For a business assistant that must be factual, it’s exhausting.
  • Less natural language: Claude Opus had a certain fluidity. GPT sounds more “machine-like,” especially in French.
  • Following complex instructions: Claude better respected multi-level constraints.

Subjective as it is, for our use case (business agents in production), GPT-5.4 was not on par with Opus 4.6 on OpenClaw.

The Core Issue: The Direct Anthropic API is Too Expensive

We could have used the Claude API directly (without OpenClaw). But Opus 4.6 pricing is prohibitive for a French Guiana agency:

  • $15 per 1M tokens input
  • $75 per 1M tokens output

With our agentic workflows (easily 100-500k tokens per client session), this becomes $100-300/day just in API costs. Multiplied by 5-10 active clients, it’s several thousand euros per month. Not viable for our setup.

Where We Stand Today

We pivoted to Claude Code (Anthropic's official CLI). It's what we used before OpenClaw, and we returned to it. It’s robust, but a step back in flexibility compared to OpenClaw.

Our current AI stack:

  • Dev/code: Claude Code (Anthropic official)
  • Simple tasks: Qwen 35B local on Mac mini
  • Client-facing: on hold, waiting for better options
  • Memory/RAG: custom pipeline based on our openclaw-memoria experience

Our Hope for Late 2026

Despite everything, I remain optimistic. Three reasons:

1. Open source is catching up with Claude Qwen, DeepSeek, Llama 4 are advancing quickly. By late 2026, we should have a local model comparable to Opus 4.6 for most tasks. No need for cloud API.

2. OpenAI will fix its weaknesses GPT-5.4 has solid foundations. Weaknesses in tooling and verbosity are fixable. GPT-6 or equivalent should close the gap with Claude.

3. Models will specialize We don’t all want a “super generalist model.” For a WhatsApp admin assistant, a well fine-tuned 7-14B parameter model for business tasks is probably more efficient than a 200B generalist.

What We Take Away, Concretely

  • Never depend 100% on an external API. Anthropic cut OpenClaw within 2 weeks. Tomorrow it could be your provider.
  • Local models are viable in 2026. Qwen 35B on Mac mini does the job for 60% of business tasks.
  • Tooling/agentic architecture remains the true differentiator. An average model with good tooling beats a large model without tooling.
  • The market will stabilize. 2027 will be less chaotic than 2026, promise (or not).

Our Promise at Primo Studio

We keep this experience to serve our clients in French Guiana. If you want to integrate AI into your business (SME, community, e-commerce), we now know:

  • Which models to choose by use case
  • How to avoid dependency on a single provider
  • How to deploy locally for sensitive data
  • How to structure an agent with real tooling

Want to explore AI for your business? Contact us for an initial conversation. We’ll honestly tell you what works and what doesn’t.

---

Neto Pompeu, CEO of Primo Studio, Saint-Laurent-du-Maroni, French Guiana