LLMs are improving fast, but the developer experience around them often isn’t. Inspired by Simon Willison’s “Better models, worse tools,” here’s a focused playbook to keep shipping reliably as models evolve. Source: simonwillison.net.
What this means for builders
Model quality keeps rising, but tooling—evals, observability, safety, and versioning—lags. The result: flaky behavior, hard-to-reproduce wins, and fragile launches.
Symptoms of “worse tools”
- Eval drift: prompts that worked on model A fail on model B or a silent point release.
- Poor observability: minimal tracing for prompts, tool calls, and retrieval steps.
- Safety brittleness: prompt injection and over-privileged tools cause unpredictable side effects.
- SDK fragmentation and vendor lock-in: switching models breaks interfaces and tests.
- Opaque RAG: missing data contracts, non-deterministic chunking, and hidden index changes.
- Unmanaged costs and latency: no budgets, no canaries, and no rollbacks on regressions.
What to do now (practical fixes)
- Treat prompts as code: version them, peer review, and write unit tests against golden datasets.
- Build an eval harness: automate regression checks across models, temperatures, and seeds before deployment.
- Instrument everything: log inputs/outputs, tool invocations, retrieval sources, and latency with trace IDs.
- Harden tool-use: least-privilege scopes, strict schemas, timeouts, allowlists, and dry-run modes.
- Stabilize your interface: use a model-agnostic adapter layer so swapping vendors doesn’t break apps.
- RAG data contracts: document chunking, embeddings, refresh cadence, and measure grounding with citation checks.
- Ship safely: canary new models, set SLOs for quality/cost/latency, and keep rollback paths ready.
What to watch next
- Structured outputs and robust function calling that reduce parsing errors.
- Stronger, standardized LLM eval suites integrated into CI.
- Vendor transparency: model versioning, change logs, and stability guarantees.
- Safer multi-agent and tool-use patterns with built-in guardrails.
Takeaway
Don’t wait for perfect tooling—create a thin, testable layer around your models, log everything, and treat prompts plus RAG like production code.
Sources
• Simon Willison: Better models, worse tools
• NIST: AI Risk Management Framework
Enjoy content like this? Subscribe to our newsletter for weekly, no-fluff AI insights: theainuggets.com/newsletter
