METHOD

How this public research initiative does the work.

Raising Agents is Adrian's personal public research initiative in cohabitation with AI agents. It is not an official Zartis publication unless explicitly marked. Adrian defines the questions, designs the experiments, owns the verdicts; collaborators are credited where they contribute. AI agents do execution work that would otherwise require a larger research group: literature sweeps, citation chaining, contract authoring, trace normalization, paper drafting passes, vendor-doc auditing. Every load-bearing artifact — the papers, the benchmark, the workbench, the bi-weekly publication — carries a citation to a primary source. The methodology is what makes the output volume defensible.

This page documents the public publication model directly. It is not a description of client delivery methodology, confidential Zartis operating process, or private engagement work. It is the answer to the reasonable question: how does one person, with collaborators and agents, produce a multi-study empirical paper, an 8,500-word position whitepaper, two hyperresearch reports across 60+ primary sources, an open-source workbench, a benchmark specification, and a bi-weekly publication — without overclaiming?

The operating model

One human. Many agents. Strict division of labor.

The work decomposes into seven stages. Each stage names what the human does, what the agent does, and what they do together. The pattern is consistent: the human owns the question, the verdict, and the load-bearing judgment; the agent owns the execution at scale.

/ 01

Question framing

The research question, the wedge claim, the scope. What is being asked and why. This is where the brand is decided: what counts as evidence, what counts as overclaim, what counts as a load-bearing concept worth a name.

Human: Owns the question. The thesis sentence. The scope of the claim.
Agent: Drafts framings, surfaces adjacent prior art, flags ambiguity.

/ 02

Literature work

Six-literature surveys, citation chains, source tensions, contradiction graphs. The R1 and R2 hyperresearch reports behind the whitepaper were each ~50 primary sources fetched, read, cross-referenced, and synthesized through a 16-step pipeline with four independent critic agents (width, depth, dialectic, instruction). The human reads the digest; the pipeline reads the sources.

Human: Selects the question. Reads the synthesized report. Owns the verdict.
Agent: Hyperresearch pipeline: width sweep, depth investigation, contradiction graph, evidence digest, four-critic adversarial review, patcher, polish.

/ 03

Experiment design

Preregistered protocols. Sample sizes pre-committed. Acceptance gates locked before runs. Paper 1's PAPER1_PROTOCOL_FREEZE.json was authored before Study 1 began executing. The human owns the design; the agent stress-tests it for ambiguity and missing controls.

Human: Hypothesis, sample size, acceptance gates, scope.
Agent: Pre-mortem on the protocol. Surfaces ambiguity. Drafts the preregistration JSON.

/ 04

Trace + contract authoring

Agent execution traces are normalized into the agent-behavior-trace/v0 schema. Behavior contracts target consequential trajectory events. Each contract is a C = (P, I, G, R) tuple: preconditions, hard invariants, governance policies, recovery mechanisms. The human authors the first contract; the agent generates variants and adversarial cases.

Human: First contract per consequential event. Decides what counts as a violation.
Agent: Trace normalization. Contract-variant generation. Workbench CLI runs at scale.
Together: Iteration until the contract catches the failure and tolerates the variation.

/ 05

Execution

3,797 runs across four domains for Paper 1. 60,000 trajectories in the Bjarnason replication corpus the whitepaper cites. The workbench's abw gate CLI emits PASS / WARN / FAIL / INCONCLUSIVE verdicts against the locked contract. The human does not execute runs by hand. The agent does not invent verdicts — it reports them.

Human: Lock the protocol. Authorize the runs. Decline to interpret before execution completes.
Agent: Run the experiments. Capture traces. Apply the locked contract. Emit verdicts.

/ 06

Drafting

Papers and the bi-weekly publication are drafted in passes. The human writes the thesis, the wedge, the closing. The agent generates body prose against the citation graph the human has approved. The human passes through three times: structural, tonal, evidentiary. Final pass is always human.

Human: Thesis, voice, closing. Final pass. The decision to ship.
Agent: Body prose, citation injection, structural drafts.
Together: Adversarial pass: the agent argues against the draft; the human re-grounds.

/ 07

Verdict + publicationthe human decides. always.

The verdict is human. Whether the evidence supports the claim. Whether the language overclaims. Whether the artifact is ready. The agent does not decide what ships. The agent helps the human verify their own draft against the standards the human has set.

Human: Verdict. Publication decision. Public-facing claim.
Agent: Pre-publication review against the brand's anti-claim list, citation completeness, voice consistency.

Operating principles

Four constraints that make the model work.

/ 01 · The human owns the verdict

No artifact ships without a human verdict pass.agents drift fast without this. i've seen it happen.

The agent can draft, audit, cite, and stress-test. The agent cannot decide that a claim is supported or that a piece is ready. This is the discipline that prevents the operating model from drifting into "AI wrote a paper that no one stood behind."

/ 02 · Every claim carries a citation

Load-bearing claims are evidence-bound by construction.

The whitepaper's "five-tradition convergence" is not an observation the human invented in conversation. It is the synthesis output of two hyperresearch runs over ~50 primary sources, with each tradition's claim grounded in arXiv IDs. The agent enforces this discipline by refusing to surface uncited claims in drafts the human will read.

/ 03 · Preregister before executing

Protocols are locked before runs begin.locks the goalposts. then the data speaks on its own terms.

Paper 1's PAPER1_PROTOCOL_FREEZE.json commits the sample size, the acceptance gates, the redaction policy, and the venue-quality bar before Study 1 starts. This is what allows the human to trust the agent's execution: the goalposts cannot move during the run.

/ 04 · Open-source the instrumentMIT license is a claim. not a footnote.

The workbench is the reproducibility surface.

Agent Behavior Workbench is MIT-licensed. The four versioned schemas (agent-behavior-trace/v0, agent-behavior-contract/v0, regression-of-the-week/v0, lab-board/v0) live at stable URLs under raisingagents.is/system/schema/*. Every claim in Paper 1 reproduces from the locked protocol files and the workbench corpus. The methodology is auditable, not assertable.

What this is not

Three honest exclusions.

Not "AI wrote the papers." AI agents executed thousands of runs, surveyed dozens of primary sources, drafted body prose against approved citation graphs, and stress-tested arguments adversarially. The thesis, the verdict, the publication decision, and the voice are human.
Not "the agent is the researcher." The agent is the research instrument. Distinction matters. A telescope is not the astronomer. A workbench is not the engineer.
Not a generalized claim about AI labor. This page documents how this lab works, not how AI agent collaboration should work everywhere. The model fits research-heavy, evidence-bound, narrow-scope work where preregistration and citation discipline are tractable. It does not generalize uncritically to creative work, judgment-heavy product decisions, or open-ended exploration.

Where this comes from

The era named honestly.

The Innovation-Lab-with-AIs operating model is not unique to Raising Agents. Apollo Research, METR, Anthropic alignment work, and the cohabitation literature converging in 2025–2026 all describe variations of this pattern: small human teams with agent-mediated execution, producing research output volumes that would have required twenty-person groups five years ago.

What this lab adds is a discipline of artifact-first publication: every claim has a workbench-reproducible artifact, every paper has a preregistered protocol, every position is backed by a literature audit, every vendor framing is checked against verbatim quotes. The discipline is what makes the operating model honest. Without it, the model produces volume without rigor.

The methodology is the product because the methodology is what the lab is teaching.

The team route through Zartis, the bi-weekly publication, and the workbench together are not separate offerings — they are the same thing surfaced at three timescales: applied adoption (teams), regular practice (publication), and reproducible instrumentation (workbench).

Surfaces of this lab

/ behavior lab

Behavior Lab

The workshop board. Concrete demonstrations on canonical tasks.
/ papers

Papers

Paper 1 (empirical lemma) and Paper 2 (whitepaper) with R1 + R2 companion reports.
/ workbench

Agent Behavior Workbench

The open-source instrument. TypeScript + Python. MIT-licensed.
/ behavior watch

Behavior Watch

Bi-weekly publication. One field note, one behavior pattern, one practical boundary. The continuous-practice surface.

Advisory: adrian@raisingagents.is. Zartis-routed team work: adrian.sanchez@zartis.com. Do not send confidential client, company, customer, or personal material through public forms.