Deep-dives on agentic systems, generative computing, and the ideas behind Mellea.
Mellea v0.8.0 adds audio as an input modality, lets you retrieve the files your sandboxed code produced, cuts m CLI startup from 3.2s to 0.1s, and rebuilds streaming as a plain async iterator. There are breaking changes to review if you use streaming or telemetry.
TypeSafe's Jev answers typed questions with probabilities instead of generating text. Mellea has been going at validation from the same direction, starting with the cheapest check that will do the job.
Vision models return prose. This post shows how to get a typed Python object back instead, using Mellea's format= parameter and ImageBlock.
Mellea v0.7.0 ships a sandboxed code interpreter, a shell tool, and a library of executable requirements, plus context compaction and plugin-based telemetry, so agents can run code and stay grounded.
Trace generation, validation, and sampling in detail. Built-in plugins reveal model calls, requirement failures, repair events, and loop iterations—all without boilerplate.
At the IBM Booth at THINK 2026, we ran a demo using every model family from the Granite 4.1 release
Agents are just programs, patterns of control flow around generative AI. So why the 80–90% failure rate? Because they're built out of prompts, not code. Mellea is a different approach.
With Granite Switch, adding validation to a Mellea program — checking that an answer is grounded, that a requirement is met, that nothing in the response was hallucinated — is a single function call against the backend you're already using. One checkpoint, a dozen drop-in validations, no second pipeline to stand up.
The industry just spent a fortnight agreeing you should write loops, not prompts. Everyone also agrees on the catch: a loop is only as good as the gate that can fail its work.
Wiring functional tests into Mellea's Instruct-Validate-Repair loop nearly doubled functional correctness on the Qiskit Human Eval benchmark, on top of what static validation already provided.
Mellea now supports MCP server tools. Discover any MCP server's tools and call them directly from a Mellea agent.
Small open-weight models can handle production-shaped work when the harness decomposes the task, validates outputs, and routes each step to the right local model.
Add semantic validation and quality guarantees to DSPy programs with Mellea's integration for structured prompting and runtime verification.
Mellea brings structured validation and automatic repair to CrewAI multi-agent systems through the instruct-validate-repair pattern.
Route most requests to a small model and escalate only hard cases to a larger one — Mellea's SOFAISamplingStrategy makes the dual-model pattern a one-line strategy swap.
Learn how Mellea's generative programming patterns add structured validation, automatic retry, and inference-time scaling to LangChain applications.
Install uv, pull a local model with Ollama, and build your first Mellea pipeline from scratch — no API key, no cloud, fully private.
How Mellea brings structured validation and automatic retry to LangChain, CrewAI, and DSPy
Use mellea's provider-agnostic backend abstraction to build LLM applications that automatically survive outages through three layers of failover: validation retries, capability escalation (SOFAI), and infrastructure switching across providers.
How we used Mellea's Instruct-Validate-Repair pattern with flake8-qiskit-migration to automatically catch and fix deprecated Qiskit APIs in LLM-generated code.
Hooks are a simple but powerful way to tap into your LLM application's lifecycle and add custom behavior without touching your core logic.
What is “generative computing,” and is it different enough from other things to deserve a name?
Presenting an opinionated view on generative AI.
