How AI agents actually operate at Thanos Labs

The team model is a small human core leading the work, AI employees carrying defined roles, & human freelancers covering specialist work such as manual accessibility audits. On public materials the AI agents are labeled AI employees, so the structure never pretends they are human.

Thanos Labs sells supervised AI agents for websites, search visibility, accessibility, cybersecurity, reporting, & customer experience. We run that system on our own site first, in public, before we ask a client to trust it on theirs.

That choice keeps us honest in three ways. Every workflow has to survive real pages, real deadlines, & real search data, not a demonstration. Every weakness shows up on our site before it can show up on a client site. & every result quoted in this case study can be pointed to in a dated report, a screenshot, or a workflow document.

This is a single case study of one production system on one website, studied from September 27 to October 13, 2026, the first weeks after the domain was registered. Evidence comes from four sources: the dated workflow & AI agent specification documents that govern the system, the audit reports the AI agents produced about the site itself, the verification crawls run after changes were published, & screenshots captured at the time a result was observed.

Three limits apply to everything that follows. A single site cannot show what will happen on a different site in a different market. Early results on a new domain are a starting baseline, not a growth record. & automated scans, however useful, are not conformance audits, a distinction we keep wherever scan results appear.

Every change to ThanosLab.com follows the same pipeline, whoever or whatever drafted it: the Web Content AI Agent builds, the QA AI Agent verifies, a human approves, the human publishes, & the live result is monitored. A passing build is not done. Nothing publishes without a QA pass verdict & explicit human approval.

• 1. Collect evidenceStart with an observation or a test result: a crawl, a scan, a screenshot, or a report. Not an opinion.
• 2. Propose a changeThe AI agent states what should change, why, and what evidence supports it.
• 3. Test the previewThe change is built on a preview and checked there, never pushed straight to production.
• 4. Obtain the required approvalA human reviews the exact change list, called the pre publish changelog, and approves or sends it back.
• 5. DeployThe human lead publishes production. AI Agents prepare and verify, they do not publish.
• 6. VerifyThe live site is rechecked after publish, from the visitor point of view, and the result is recorded.

Source: Thanos AI Agent Workflows, governing rule recorded October 5, 2026.

The more independently an AI agent can act, the heavier its testing, approval gates, and audit requirements. That principle was adopted on October 13, 2026 & is now written into every AI agent spec.

Most AI agents sit in the advisory tier on purpose. Autonomy is earned per job, not granted per AI agent, & a tier can move only when the testing behind it supports the move.

• Known answer testsEach AI agent carries 30 to 50 known answer test cases, run before it is trusted on a new platform or a new kind of work.
• Tagged knowledgeEvery knowledge document is tagged with its source, its date, and its review date. Primary sources are preferred.
• Approvals feed memoryWhen a human approves, edits, or rejects AI agent work, that decision feeds back into AI agent memory, so the same correction is not needed twice.
• Focused retrievalAI Agents pull the specific record they need, not a data dump, and they summarize third party material in their own words instead of copying it.
• Strict scoringIn evaluation, an invented source counts as a full failure, in the same class as an invented fact. Answering with low confidence when the AI agent should have held back also fails, and so does holding back needlessly.

On top of AI agent evaluation sits the publish QA battery for ThanosLab.com. A preview must pass lint & build checks, HTML validation, a broken link check, an accessibility scan with axe core against WCAG 2.2 AA tags with zero violations as the gate, copy checks, SEO head checks, image checks, an AI crawler reachability check, a click through smoke test, & a Copyscape originality check before publishing. If any gate fails, the work goes back with the failing check, the evidence, & fix guidance, & the full battery runs again after the fix. There are no partial re runs.

Documents get their own battery. Before a PDF ships, it is checked with a screen reader traversal simulation, veraPDF validation, contrast checks, & a proof that no text is left untagged. That is an automated battery, not a claim of real screen reader testing. A human spot check with VoiceOver remains part of the process.

This is one complete pass through the pipeline, with dates & numbers that can be checked in the report itself.

• EvidenceAI Agents crawled all 52 live pages on ThanosLab.com and combined the crawl with PageSpeed data and read only Search Console data. The audit scored the site 78 out of 100 overall.
• ProposalThe clearest fix was metadata: 19 meta descriptions were over length and 17 titles were over length.
• Preview & gatesThe fix was built, checked with type, lint, build, and SEO head gates, and verified on a preview. The SEO head gate passed on 52 of 52 pages.
• Approval & publishThe human lead approved and published the change on the evening of October 12, 2026.
• VerificationA fresh production crawl of all 52 pages confirmed the fix live: maximum title 59 characters, maximum description 160 characters, and zero canonical errors. The updated overall score was 83 out of 100, with on page at 97 out of 100.

The same report also recorded what the fix did not solve, including the findings register summarized in Section 09. A case study that only lists wins is marketing. The register is part of the result.

The October 12 report kept a findings register, & the open items stay in this case study because they show how the system handles unwelcome news: it records it, assigns it, & keeps it visible.

• BacklinksSearch Console showed zero external backlinks at the time of the audit, and the authority score was 5 out of 100. Earning editorial backlinks is the current outreach priority, with no paid placements and no link farms.
• Tracked site findingsThe register also tracked items such as redirect discipline, sitemap dates, and a blog template mobile speed score of 70. Fixes go through the same pipeline, in priority order, and each fix is verified live before it is called done.

• They do not publishNo page, post, email, or message goes out as Thanos Labs without human approval of the exact final version.
• They do not take state changing security actions on their ownIn security work, AI agents collect read only evidence and propose. A human approves actions such as disabling an account or blocking an address.
• They do not scan beyond permissionProspect security checks are passive and external only. Anything active requires written consent first.
• They do not hold credentials in chat or in filesLogins and keys go through secure storage, and per client credentials are never reused across clients.
• They do not invent support for a claimIf the evidence is not there, the claim is not in the copy. That rule produced the limits in this case study, including the branded query note on the AI Overview result.

For practitioners. The bottleneck in agent adoption is rarely model capability. It is governance: who may approve what, what evidence a claim needs, & what happens when a check fails. A small team can operate a multi AI agent system safely when those answers are written down before the AI agents run, and when approval is treated as a designed step in the workflow rather than a courtesy copy of the output.

For researchers. Three features of this case may be of interest to groups studying AI agents in organizations. First, autonomy is tiered per AI agent & tied explicitly to evaluation load, which makes scrutiny a measurable design variable. Second, human approval decisions are fed back into AI agent memory, so oversight doubles as training data for future behavior. Third, the system records negative findings about its own site in the same register as its wins, which offers a natural check on selective reporting in AI agent produced case studies, including this one.

We offer these as observations from one case, not as general laws. The value of a case study like this is that its claims can be checked, challenged, & revisited as the record grows.

What is demonstrated today. Neuralink reported 21 participants enrolled in trials worldwide as of January 28, 2026, reported by Reuters, with its first participant controlling a cursor, playing games, browsing, & posting by thought. The company reported zero serious device related adverse events, a company claim carried by Reuters, & no peer reviewed outcome paper has been published yet. Synchron reported its COMMAND feasibility study of 6 US participants over 12 months with no brain or vasculature device related serious adverse events, & in 2025 an ALS participant controlled an iPad by thought, with neural signals sent to an external decoder. Researchers at UCSF & UC Berkeley reported in Nature Neuroscience on March 31, 2025 that one participant, unable to speak for 18 years, had attempted speech streamed into voice by a deep learning decoder. Precision Neuroscience received FDA 510(k) clearance on April 17, 2025 for a surface electrode array used for up to 30 days, while its fully implantable system remains in development.

What the TEDx record shows. Our October 2026 survey of 20 relevant TEDx talks, drawn from official talk descriptions and transcript summaries, splits into two separate conversations. Implant talks are medical & restorative: Thomas Oxley at TEDxSydney presents a stent based interface as a direct brain to AI link for people with tetraplegia, & Polina Anikeeva at TEDxCambridge focuses on restoring function. Personal agent talks are external software: Rohit Prasad at TEDxBoston describes a lifelong personal AI assistant, & Mei Lin at TEDxUTA describes a career digital agent, with no implant in either model. Critical voices qualify the direction further, including Anikeeva arguing against uploading the brain, Heather Berlin at TEDxKC on consciousness, Allan McCay at TEDxSydneySalon on human rights, & Dave Segal at TEDxStGeorge presenting a non invasive earbud as an alternative to implants. No surveyed TEDx talk joins the two halves. The join is the thesis. For accuracy, we also note that the talks by Ray Kurzweil, Tom Oxley, Tom Gruber, & Nita Farahany most often quoted on this topic are main TED talks, not TEDx, & are not cited here as TEDx.

Why the interface wording matters. In every demonstrated system, the implant is a sensor. Decoding, speech synthesis, applications, & any agent reasoning run on an external computer, phone, or cloud system, & basic constraints point the same way: an implant has strict limits on volume, heat, & power, while a personal agent needs a large model, long term memory, tools, & frequent updates. The defensible claim is therefore an implanted interface to a personal agent, not an implant that is the agent.

What remains open. Long term durability over five to ten years is unknown. No implanted motor interface has full FDA approval as a commercial product. Public acceptance for healthy use is weak: Pew research found about 69 percent of adults worried versus 34 percent enthusiastic about a brain chip for a healthy person, & 66 percent said they would not want one. UNESCO adopted its first global recommendation on the ethics of neurotechnology in November 2025, treating neural data as sensitive & mental privacy as a right to protect.

Why this thesis belongs in an operating case study. If the interface between a person and their agent keeps moving closer, from screen to voice to, one day, perhaps a neural channel, the governance questions in this case study become more important, not less. Who approves an action, what evidence supports it, what autonomy tier the agent holds, and how the most personal data a person has is protected are design decisions to make now, while the agent still lives on a phone and a website, where the answers can be tested in public.

A client gets the same pipeline, pointed at their site: an AI agent team that works continuously, an easy to read dashboard & reports that lead with the key takeaway, and a human approval gate before anything public changes. The work is visible in dated reports, not described in vague promises.

The honest trade off is also part of the model. Human gates make some steps slower than a fully automated tool. We accept that trade off, because the alternative is fast work that nobody checked.