Three active projects

Evolving Agents Labs is an open-source lab, and three of its projects are active. lora-kernel serves small local models as trained specialists. agentvcs versions and merges the harness those models run in, and attributes each change in a result to the change that caused it. gene-evidence writes evidence reports on predicted genes in which every sentence cites its source. The idea they share: what an AI system answers, changes or reports should be traceable to what produced it. One connection is built today; two are planned — how they connect.

lora-kernel

Can one small local model, given a trained specialist per job and a library of notes it has learned to navigate, do an organisation’s routine work — and hand the rest to a frontier model?

active · flagship An operating layer for specialized local AI agents: one resident model, a LoRA per job, facts in notes a person can edit, a router that abstains.

Built: specialists that beat the bare base on their job, a router that serves no foreign text locally, edits answered without retraining — measured. Not yet: no installable package, no real traffic.

Why it matters: if it works, an organisation’s routine AI work is served from one small local GPU — private, cheap to run, its facts in notes a person can read and edit — and only the rest is sent to a frontier model. The saving is not measured yet.

agentvcs

When an agent’s harness changes while it runs, can every result still be traced to the change that caused it — and can two diverging harnesses be merged on the evidence?

active · Rust core v0.1 Version control for agent harnesses: every step stamped with the harness version that produced it, patches gated, blame per metric, and a three-way merge whose real conflicts go to Claude Code — its only LLM.

Built: the Rust core, CLI, MCP server and Python SDK; attribution on a real model (below). Not yet: no PyPI package; no real Claude Code merge session recorded in the repository.

Why it matters: harnesses now change while agents run — a prompt, a model, a router. If it works, you know which change moved which result, and two harnesses that diverged are merged on evidence rather than by guess. Attribution has run on one real model.

gene-evidence

Can every sentence of a report on a predicted gene be traced to the tool output it rests on?

prototype · stage 1 Auditable evidence reports for gene predictions the reference annotation does not have: an evidence graph of tool outputs, and a report that cites it in every sentence.

Built: stage 1, deterministic, no language model; a development run on rat. Not yet: a model that drafts the report. Known limitation: it explains a candidate; it does not pick which to validate first.

Why it matters: a wet-lab experiment on a candidate gene costs real money. If it works, every claim in the evidence report can be checked against its source before that money is spent. The checkable report is built; choosing what to test is not.

Three workbenches in one long room. On the left a card-catalogue cabinet, three of its drawers lit by a small brass radar dish; in the middle a brass machine whose paper ledger carries wax seals that change colour at a swapped gear; on the right a naturalist's bench with a printed report tied by red threads to pinned evidence cards. A solid green cord joins the first two benches; two dashed ochre chalk lines lead from the third bench to the other two. — Three benches, one room: lora-kernel, agentvcs, gene-evidence. The solid cord exists today; the dashed lines are planned.

How the three connect

What they share is one rule: a statement that does not cite its source does not ship. gene-evidence’s check-report fails a sentence whose citation does not resolve; lora-kernel’s endpoint withholds an answer whose citation fails the referee’s check.

Solid green: exists and runs today. Dashed orange: a plan, not built. gene-evidence uses neither project yet.

connection	status	what it is
agentvcs ↔ lora-kernel	built	agentvcs’s lora-kernel example turns nine measured lora-kernel configuration changes into gated patches and hands a branch conflict to Claude Code. Imported, not live: nothing is re-run.
gene-evidence → lora-kernel	next — not built	A lora-kernel expert that drafts or critiques the report from the evidence graph, gated by gene-evidence’s citation checker. Today gene-evidence runs no model.
gene-evidence → agentvcs	next — not built	Each evidence run as an agentvcs run, so a database release or rule change is a patch and `blame` traces a change in the report to it. Today the versions are recorded in `versions.json`.

The flagship · lora-kernel · Apache 2.0

The LoRA is not the textbook. It is the specialist who knows how to use the library.

One GPU, one small resident model, and a shelf of QLoRA adapters — each an expert defined by its corpus. An OpenAI-compatible API serves each request with the matching expert and sends everything else to a frontier model. Facts live in short notes a person can read and correct; the adapter holds the habit of navigating them.

A reasoning expert served the way its corpus taught it: 90/90, against 11/90 through tool_calls (79 : 0 paired). Same adapter, same cases.
Trained on walks over real regulations, it cites on a family it never saw: 18/23 against the untrained base’s 9/23.
A number edited in the notes after training: 17/17 rows answer the new value, 0 stale.
The default router serves foreign text locally 0/600 times, against the keyword dictionary’s 294/600.
Inbox triage on a new base: 471/475, a tie with its earlier release.

What is measured, and what is not → The architecture article Repository

A reader at a desk before a wall of card-catalogue drawers labelled Procedures and Encyclopedia; a small radar dish on the desk lights three drawers; through an open door, a domed building on a hill labelled frontier. — The specialist, the library, and through the door the frontier for everything else.

The second project · agentvcs · Apache 2.0

agentvcs

Change your harness while it runs and still know what worked.

Every step of a run is stamped with the harness version that produced it, so a patch applied mid-run can be gated, attributed with blame, rolled back or resumed from a checkpoint. Diverged harnesses merge dimension by dimension; each real conflict goes, with its evidence, to Claude Code, agentvcs’s only LLM, and the resolution is checked before it is recorded.

Repository → The July 2026 write-up (earlier Python version)

Measured, and not

A Rust core, CLI, MCP server, Python SDK and a conformance suite.

ran On Qwen2.5-1.5B on llama.cpp, a patch applied mid-run passed a paired gate with a recall gain of +0.444, and blame attributed +0.442 to exactly one patch — the one applied (RUN_REAL.md).

not built No PyPI package; no real Claude Code merge session recorded; no bisect --exec, remotes or sync.

A brass machine on a bench with one freshly replaced gear; a paper ledger unrolls from it, each line sealed in wax, blue before the change and orange after, and a red ribbon ties the new gear to the first orange seal. — Every step stamped with the harness version that produced it.

The newest project · gene-evidence · Apache 2.0

gene-evidence

Auditable evidence for gene predictions the reference annotation does not have.

Some predicted genes are absent from the species’ reference annotation: missed genes, or artefacts. For each, gene-evidence analyze builds an evidence graph — locus context, retrocopy signature, Swiss-Prot homology, Pfam domains, ORF sanity — each node one tool output with its version and parameters. The report cites nodes in every sentence and check-report fails if one does not. No language model; the rules are in EVIDENCE.md.

Repository → The rat run

The first run, on rat

A development run, not a test of ranking quality.

947 candidates; 578 of 947 (61 %) sit on loci the reference calls pseudogenes, which is why locus context is read before homology. Two runs gave byte-identical graph and report, and every sentence passed the citation check; 35 unit tests. The top 100 are all tier 1 and repeat-rich families fill them, as written down before the run: 32 hit the Smok kinase family, 24 T-cell receptor or immunoglobulin variable segments, 2 retroviral Pol polyproteins (rank 1 among them).

known limitation On plant genomes, against long-read transcript evidence, the tiers ranked supported candidates worse than the predictor’s own confidence score. Use the report to explain a candidate, not to choose which to validate first (README).

Bar chart of the 947 candidates of the rat development run by what the reference annotates at the locus: pseudogene 578, nothing 285, protein-coding gene on the other strand or in an intron 45, non-coding gene 39. — The rat run by what the reference annotates at each candidate’s locus.

One candidate's section of a gene-evidence report: evidence for, evidence against, context and a suggested validation experiment, each line ending in bracketed graph-node citations. — Rank 1 of the rat run: every line ends in the graph nodes it cites.

The other public repositories

evolving-agentspaused The lab’s monorepo: a Claude Agent SDK plugin, the ai-os subtree, the memory package and the 2025 toolkit.
physics-verifierspaused Physics tasks with exact truth for RL rewards; the reward-hacking claim did not hold, the generator stands.
nightshiftpaused A memory-engine proof of concept for Claude Code; nobody has measured that it helps.
evolvingagentslabs.github.ioactive This website and its nightly number check.

Twenty-five more are archived and read-only, listed in the archive. agentvcs is among the projects that passed the ANFAIA Research Lab.

Get involved

Questions, reproductions that disagree with a number here, and proposals are welcome as issues: lora-kernel · agentvcs · gene-evidence · everything else under github.com/EvolvingAgentsLabs.