Towards AI’s AI coding playbook: 3 workflows that actually ship
There is a gap between AI agents that write code and AI agents that write code you can ship. The Towards AI team published their internal playbook for closing it.
Three coding workflows for shipping with agents
The full article covers three AI coding workflows plus the context rules, debugging habits, tests, and review guardrails the Towards AI engineering team uses in production. It is aimed at developers who have moved past casual prompting and want a repeatable system, not just working demos.
⚙️ The five-run rule for LLM testing
Towards AI co-founder Louis-François Bouchard shared a practical testing heuristic worth keeping: for stochastic LLM tests, run the same case five times before trusting the result.
The math behind it is concrete. A test that passes 90% of the time looks fine on a single run. Run it five times and there is roughly a 41% chance you will see at least one failure. Record the pass rate across all runs and whether it passed all five. Keep the prompt, model, temperature, and source context fixed across runs so you are measuring the test, not changing it.
️ Live workshop: personal AI engineering setup
Bouchard is running a 90-minute live workshop with Packt on September 8, 2026 (8:30 PM to 10 PM GMT+5) covering how he structures his own agent setup around Claude Code and Codex. Topics include syncing skills and knowledge across agents, turning repeated corrections into reusable skills, splitting work between Claude Code and Codex, managing usage and token limits, and using an open-source vault template. The workshop targets developers already using coding agents who want a more structured system around them. Registration is on Eventbrite.
This week’s reading list
- How The Fast Fourier Transform Actually Works: builds FFT intuition from sine waves through a Python DFT implementation, covering aliasing, spectral leakage, and windowing.
- Azure Storage for AI Inference: covers Microsoft Foundry’s three-level prompt caching, KV cache offloading beyond GPU memory with NIXL and LMCache, and model-loading optimizations.
- Your KV Cache Is Bigger Than Your Model: on gpt-oss-120b across two H100s, capacity drops from 504 concurrent conversations at 4K tokens to 15 at 128K tokens. Six techniques for recovering capacity are covered.
- Watermarking Text Generation Efficiently: explains token-probability watermarking including Google’s SynthID-Text and where paraphrasing, translation, and tokenization break the signal.
- Beyond RAG: Agentic Memory Architectures: separates episodic, semantic, and procedural memory and adds consolidation and reflection layers. Compares approaches with Mem0, Letta, and LangGraph.
I’m founder of BizStack at Cagri Sarigoz LLC and a passionate advocate for entrepreneurs.
With over 14 years in tech, marketing, and AI, including my role as Head of SEO at CitizenShipper and co-founder of TaleBot at Intale AI, I’m dedicated to sharing genuine, useful product insights and tips.
At BizStack, I aim to cut through the digital noise to provide clear, actionable advice.
And more than all else, I’m a father to a (always) little girl and a husband.
Contact me at [email protected] for assistance.
