A personal public research initiative for delegation-grade agents.

Raising Agents is artifact-first, evidence-bound, and anti-hype. It is a personal research initiative by Adrian Sanchez de la Sierra, developed alongside his AI and innovation work at Zartis. It is independently written, while many of the experiments and infrastructure are made possible by Zartis' investment in applied AI research. It is not an official Zartis publication unless explicitly marked. The work asks why adoption is high but value is still scarce: agents can produce plausible work, but significant work is not safe to delegate until authority, evidence, thresholds, receipts, and review are explicit.

What this isnewsletter is how it travels. program is what it is.

Two public loops, one evidence shelf, one private route:

/ behavior watch

Behavior Watch

The public field journal. One failure pattern, one public-safe artifact, one delegation boundary. Selected experiments and public-safe reconstructions travel here. Subscribe →
/ free experiments

Free Experiments

An agent-led intake for serious agent behavior problems. Selected public-safe ideas may become free Lab experiments or Behavior Watch case files. Talk to the lab →
/ evidence shelf

Research, Lab, and Workbench

Papers, reproducible experiments, schemas, and open-source tooling for the parts of the delegation stack that can be published.
/ work with us

Advisory and Zartis route

Start with Adrian for advisory and framing. Larger, private, implementation-heavy, or team-specific work routes through Zartis.

If a problem can be public-safe, send it through Free Experiments. If it depends on private material or production specifics, use Work With Us.

What this is notevery one of these was a temptation we said no to

Not client disclosure. Private client work, Zartis-confidential material, credentials, raw transcripts, and engagement-specific detail do not appear here.
Not a hosted SaaS. The workbench is open-source, local-first. A product path may follow; today there is no SaaS to sell.
Not a general AI publication. The scope is narrow on purpose.
Not an opinion outlet. Every claim has an artifact behind it, or it does not ship.

Authorship and route

Adrian leads the public voice; contributors are credited where they work.

Raising Agents is a personal public research and publication initiative by Adrian Sanchez de la Sierra. Adrian leads the editorial voice, research framing, and publication decisions. Genesis Rojas and Radu Simonescu contribute to the research and lab work where credited. Zartis supports the lab through its investment in applied AI research and is Adrian's professional home; it is also the route for larger teams that want to apply the work in production.

Views and editorial choices are Adrian's unless a page is explicitly marked as official Zartis material. Advisory can start with Adrian; client work, workshops, assessments, implementation support, and scaled delivery route through Zartis.

The Lab outputs are the result of systematic behavioral research.legally and editorially clean. all of it. Public artifacts use public, synthetic, redacted, or explicitly approved material. Paper 1's executed studies run on the open Agent Behavior Workbench; the corpus is reproducible from the locked protocol files. The conceptual framing is deliberate: rigorous evidence over hype.

Advisory: adrian@raisingagents.is. Zartis-routed team work: adrian.sanchez@zartis.com.

How this work is funded

The papers, workbench, benchmark, Behavior Watch, and selected Free Experiments are public research and publication work. Advisory may start with Adrian; Zartis-routed workshops, team assessments, and implementation support are the professional path. The choice to make evidence and tooling public while protecting private adoption work is deliberate: the lesson can be public, the private work should stay private.

No external funding, no vendor sponsorships, no client material. If that changes, this page changes first.

Boundaries

The lab does not publish private client material, confidential Zartis IP, private chain-of-thought, raw transcripts, or any content that would compromise an engagement, colleague, company, or user. The benchmark and workbench corpora are public, synthetic, or redacted. The papers cite locked protocol freezes and reproducibility packages. The boundary between public research and client work is explicit by design.

Reader submissions should follow the same rule: do not send confidential company, client, customer, or personal data. Use public, synthetic, or generalized examples.

Open public work

Home, Watch, Lab, Papers, Workbench, Bench, and public schemas. These can use public sources, synthetic examples, redacted examples, approved artifacts, and open-source tools.

Advisory and Zartis-routed work

Advisory, teams, workshops, assessments, implementation support, and applied adoption. These conversations do not become public case-file disclosure.

Blocked from publication

Client traces, client names, engagement-specific details, confidential Zartis IP, raw private transcripts, colleague-identifying material, credentials, and anything that would compromise a user or company.

On naming: Agent Behavior Engineering is presented as a synthesis label over a convergent research cluster, not as an established academic field. Delegation-grade agents is presented as a vision and target, not as a solved engineering recipe. Behavior contracts are one mechanism, not the whole reliability architecture. Each of these scopes is named where the concept is introduced.

Colophon

How this site is built

Source Serif 4 for body, JetBrains Mono for code and metadata, system sans for UI, Caveat for margin notes. Oxide red (#9C3A24) reserved for FAIL / regression / violation. Off-white paper, warm-black ink. Dense, evidence-first prose. Hand-built, no framework, no SaaS dependencies on the front of the house.

Source for the site: github.com/Adriansdls/raising-agents. The research substrate — Agent Behavior Workbench, OPBR-Bench, the detector suite, the statistical scripts — lives at github.com/raising-agents/agent-behavior-workbench.