Claude worked alone for 21 hours: what real AI autonomy looks like

On September 23, 2026, Anthropic announced that an agent built on Claude had spent 21.5 straight hours — 949 chained sessions, 215.6 million tokens — searching a database of 1.9 billion proteins without continuous human supervision, until it identified a previously undescribed enzyme system with DNA repeats resembling those behind CRISPR. At Evicron, an AI and custom software studio based in Barcelona, the biology isn’t what interests us here: what matters is that this is, by a wide margin, the best-documented example to date of what “autonomous agent” actually means once an AI company measures it in hours and sessions instead of a product tagline. It’s a useful yardstick for any SME being sold “an AI agent” for its business.

What the agent actually did

The system, which Anthropic calls ART (array-associated reverse transcriptases), turned up in the DNA of bacteriophages — viruses that infect bacteria — after an autonomous sweep through a protein database. According to Anthropic’s official announcement, Feng Zhang, one of the scientists who brought CRISPR into living-cell gene editing, called the work “an exciting example of how AI agents can contribute to biological discovery.” The company is upfront about the limits: the system hasn’t been shown to edit genes, and its exact biological function is still unknown. In fact, as Al Jazeera reported, several microbiologists pushed back on the framing — for some specialists, finding “CRISPR-like” systems in public databases is closer to routine genome mining than a disruptive breakthrough, and resembling CRISPR doesn’t mean it does what CRISPR does.

That pushback from the scientific community itself is as relevant to a business reader as the discovery. No one selling “autonomous agents” volunteers that caveat by default.

The lesson isn’t about biology, it’s about sustained autonomy

What changes with this case isn’t that a model “knows a lot” — it’s that an agent ran nearly a full day of specialized work, deciding on its own what to try next, with no one reviewing every intermediate step. That’s the difference between an assistant that drafts an email when you ask, and an agent you hand an open-ended goal on Monday and get a result from on Tuesday.

That distinction is exactly what gets blurred in most commercial pitches for “AI agents for business.” When a vendor talks about “autonomy,” it’s worth asking which tier their product actually sits in:

• One-off assistance: answers a question or runs a specific action each time you ask. That’s not autonomy, it’s a tool.
• Supervised automation: chains several actions together but stops for confirmation at critical points.
• Real sustained autonomy: works for hours toward a goal, decides its own next step, and only surfaces a final result for human review — the Claude/ART case.

Most of what’s currently sold as “agents” to Spanish and European SMEs sits in the first or second tier, not the third. That doesn’t make them useless — automating one concrete process already saves hours — but conflating the tiers inflates both expectations and budgets.

How to evaluate an AI agent project with this yardstick

1. Ask for the goal, not the demo. Anyone can show an agent nailing a rehearsed question. Ask what open-ended goal it can pursue unattended, and for how long.
2. Ask what happens when it goes wrong midway. A 21-hour unsupervised process can also rack up 21 hours of error if the goal was framed badly. Whoever sells you real autonomy should be able to explain how that failure gets caught and cut short.
3. Demand the result with its limits, not just the headline. Anthropic published the finding alongside what it still doesn’t know. A serious applied-AI provider does the same with your project: what works, what’s still unproven, and what’s left to validate.
4. Compare the real cost, not the license price. 215.6 million tokens for a single run is a reminder that sustained autonomy costs compute, not just a monthly subscription; any quote that doesn’t address that variable cost is incomplete.

How we approach this at Evicron

When we build applied AI agents for a client, we always start by pinning down which of those three tiers the agent actually needs to sit at — almost never full autonomy, which rarely pays off in real business processes compared to a design with human review checkpoints — and we put that in writing before touching a line of code. We already covered what to demand from an AI provider once you hand an agent real permissions; this case adds one more question to that list: ask them to name, honestly, which autonomy tier what they’re selling you actually sits at — with the same honesty Anthropic showed about the limits of its own finding.

We’ve been building custom software and applied AI for companies across twelve sectors and all of Europe since 2019, with over 200 projects delivered and no vendor lock-in. If someone has pitched you an “autonomous agent” and you’re not sure whether it’s a tool with a new name or something that can genuinely hold a goal unsupervised, get in touch: the first discovery session is free, and you’ll leave it knowing which autonomy tier your process actually needs — not the one that’s easiest to sell.