Towards AI

09/05/2026

build a retrieval eval in an afternoon.

take 30 real questions. write down the page each question should retrieve. run search and mark whether that page appears in the top 20.

slice the misses by query type: IDs, tables, dates, exact titles, and natural-language questions. each slice points to a different fix.

09/04/2026

new model benchmarks are interesting. your failing test suite is still the acceptance test.

09/04/2026

Towards AI is now an official OpenAI services partner!

Through the OpenAI partner network, we support engineering teams taking LLM applications from early prototypes into production.

We focus on:

• Designing reliable RAG pipelines and multi-agent systems
• Building custom evaluation benchmarks and fine-tuning pipelines
• Architecture reviews and hands-on engineering for scaling on OpenAI's platform

If your team needs implementation support or technical advisory, let's get in touch.

09/04/2026

pick the coding tool by where the work must continue.

- repo and terminal: agent with direct file and test access
- live UI exploration: editor or browser session
- unattended work: durable task state, logs, checkpoints, and resume

model rankings do not tell you whether the task survives a closed laptop.

09/03/2026

We recently talked with engineers around us on their biggest complain with AI coding agents

Apparently the friction with AI coding agents comes from the setup around them, not the underlying models:

Here are some of their complaints:

• Context windows fill with noise after a few iterations.
• Architectural rules and formatting preferences reset every session.
• Custom instructions stay trapped inside whichever tool you wrote them in.
• Hitting a rate limit in Claude Code or Codex stalls the whole workflow.

If this is also you, we have something for you.

Our CTO, Louis-François Bouchard, is running a free live session with Packt on building a portable AI engineering system.

He will walk through the architecture he uses to keep agents reliable:

• Building a synced Markdown vault that holds instructions across tools without vendor lock-in
• Writing skills that change agent behavior rather than storing static prompts
• Orchestrating multiple tools, using Claude Code for planning and Codex for bounded ex*****on
• Setting up feedback loops so your agents retain project preferences over time

Those who attend will also get the vault template and patterns to adapt for their own stack.

Register here:

Build a portable coding setup that remembers how you work, improves, orchestrates Claude Code and Codex, and keeps working across tools.

09/03/2026

RAG still solves a common problem: the answer depends on information that does not belong in the model weights.

09/03/2026

static and live knowledge need different paths.

policies, API maps, and schemas can sit in a stable cached prefix. tickets, prices, and recent traces should be fetched when needed.

prompt caches are prefix-sensitive. reorder the static documents and the cache may miss.

09/02/2026

measure which parts of the prompt are sent repeatedly before changing models.

09/02/2026

split prompt tokens into four buckets before changing models:

1. stable instructions
2. retrieved evidence
3. tool schemas
4. replayed tool output

the fix depends on the bucket. cache stable prefixes, retrieve evidence, load fewer tools, and store bulky results outside the window.

09/01/2026

set a time, cost, or attempt limit before an agent starts repeating work.