Evolving Agents Labs every number checked nightly

The new flagship · lora-kernel

The entire agentic system is a set of QLoRA adapters over one base model.

No router model. No JSON schemas in the system prompt. No parser guessing whether the model meant to call a tool. One GPU, one resident base, and a pool of small weight deltas swapped per request by vLLM. The multi-agent framework does not get simpler — it turns into weights.

Read the article → Architecture Repository

specified · nothing built  Every external claim is marked [read] and cited; there is no [ran] yet. Below: the eight steps it runs in.

How it works

 

 

the host — vLLM, one GPU

BASE MODEL  ·  resident

adapters — the whole agentic system

harness.lorakernel
clinical-adminexpert
contract-reviewexpert
incident-triageexpert
the passcost: —
code this referral
flag this clause
page the on-call
FRONTIER MODEL — verifies every branch

acceptance, per region

clinical—
contract—
triage—
click any step, or use ← →

What each piece becomes

harness.lora

What if the tool protocol lived in weights, not in the prompt?

Specified. One adapter trained on nothing but the execution protocol — tool syntax, state transitions, error shapes, and action tokens emitted natively rather than recovered by a parser. The schema leaves the context window. The domain adapter thinks; the kernel acts. It has to beat a training-free baseline this organisation already measured, and a constrained decoder that makes malformed output impossible rather than unlikely.

domain QLoRAs

What if an agent were a few hundred megabytes of delta?

Specified. User space. Hot-swapped per request, batched by vLLM's multi-LoRA serving, versioned like code, scored, promoted and retired. Several per sub-domain, competing.

the router

What if routing were a by-product rather than a call?

Specified. While the frontier verifies, the router is the acceptance rate — free, because the tokens were generated and the pass was already happening. After withdrawal it is a small model fitted to the surface that phase produced, and it is the only thing left where the frontier used to be.

what stays text

What deliberately never becomes a weight?

Two things, and it is not aesthetics. Memory, as markdown under git — a weight delta cannot be read, diffed, cited or corrected by a person. And execution, as a sandbox where tools actually run. Everything this lab has measured about memory says the durable asset is the part a human can read.

The engineering state is stated rather than assumed: multi-LoRA over one target and tree-structured verification ship in vLLM today; LoRA-as-drafter is an open RFC filed in August 2026, and the KV cache across different adapters is the hard part, named rather than waved at.