From Chatbots to Colleagues: How Agentic AI Is Changing Work in 2026

Agentic AI is the shift from software that answers a prompt to software that plans and carries out several steps on its own. In 2026 that shift moved out of demos and into everyday work: coding agents open pull requests, research agents search databases overnight, and operations agents file tickets and chase approvals, with people setting the goals and checking the results.

Agentic AI plans and acts, where a chatbot only replies

For the past few years, “AI” for most people meant a chatbot: type a prompt, get a paragraph back, paste it somewhere else. That is useful, but the chatbot does nothing until it is asked, and it stops when it has answered.

An agentic AI system works differently. Given a goal, it breaks the task into steps, chooses tools, acts, checks what happened and tries again. Ask a chatbot to “research competitors” and you get a summary. Ask an agent and it can find the companies, read their pricing pages, build a comparison table and draft a report for a person to review. TSN’s explainer on what AI agents are and how they differ from compound AI systems covers the architecture in more detail.

Where AI agents are already working alongside people

Agents are now being used across ordinary business functions. In HR they screen CVs, schedule interviews and draft onboarding documents. In legal teams they review contracts and flag unusual clauses. In finance they help with forecasting, fraud checks and procurement. In software development, AI coding assistants have become standard tools, and vendors are shipping agents that fix code on their own, such as GitHub’s Agentic Autofix and AI Scan.

Responsibility for these systems is also moving. Rather than sitting with small innovation teams, AI is increasingly overseen by finance, operations and compliance functions, and new roles have appeared around it, including AI governance leads, model risk analysts and people who design how humans and agents hand work to each other.

The open-source stack behind AI agents

Agents need more than a chat window. Most enterprise deployments combine several layers, many of them open source:

• vLLM and SGLang serve models at scale on GPU clusters, where latency and throughput matter.
• llama.cpp and Ollama run capable models on laptops and edge devices, so sensitive data can stay on local hardware.
• LiteLLM routes requests across model providers behind a single API, which helps teams manage cost and fall back when a provider is down.
• Unsloth makes it cheaper to fine-tune and run open models locally.

These are actively maintained projects with versioned releases, and teams increasingly mix them to suit the workload rather than relying on one stack. Small language models trained for a narrow domain are part of the same trend: on a specific task they can be good enough to replace a much larger general model, at lower cost and without sending data off site.

How agents change jobs: tasks move, judgement stays

Consider a marketing manager. A couple of years ago they might have asked a chatbot for blog ideas. With an agent, they can have it gather competitors’ recent campaigns, draft a content calendar, prepare assets for review and report on engagement, all within limits the manager sets.

Or take a supply-chain analyst who used to spend days collating supplier, port and weather data. An agent can watch those feeds continuously, flag likely delays and propose alternative routes, leaving the analyst to decide on the exceptions that need human judgement.

The pattern is consistent. Agents take on volume, speed and pattern-matching; people keep ambiguity, ethics, creativity and relationships. What disappears first is tasks rather than whole jobs. The change is not painless for people in highly procedural roles such as data entry or basic support, and the organisations that handle it well retrain and redeploy staff rather than simply cutting them.

Anthropic’s Claude agents found a CRISPR-like enzyme system in 21 hours

The clearest sign of what agents can do comes from research. On 23 September 2026 Anthropic said that roughly 950 Claude agents, working for 21 hours and using 210 million tokens, searched a large DNA database for reverse transcriptases, enzymes that copy RNA into DNA. The agents gathered more than 200,000 of them, picked out 3,500 new candidate systems and narrowed those to 20 for written reports, work Anthropic says can take an expert weeks to months.

One agent spotted a repeating DNA pattern next to an unusual reverse transcriptase gene in bacteriophages. Anthropic’s scientists then tested it in the lab and named the system array-associated reverse transcriptases (ART). Its repeat array resembles CRISPR, but Anthropic says it does not yet know what ART does, and all lab work was done by human scientists. The details are in its technical report.

Anthropic’s chief executive, Dario Amodei, has argued in his essay Machines of Loving Grace that AI could compress the progress biologists would make over 50 to 100 years into five to ten. ART does not prove that, but it shows agents doing the slow screening work that usually comes before a discovery.

Agentic AI needs identity, limits and a human sign-off

Agents also bring risks that chatbots did not. A bad chatbot answer is embarrassing; a bad action taken on someone’s behalf can move money, delete data or open a security hole. The risks are practical and present: agents can make harmful decisions, be manipulated by malicious content, or spread bias quickly through high-stakes processes.

The answer is governance built into the system rather than bolted on. Frameworks such as the NIST AI Risk Management Framework give organisations a structure for mapping and managing those risks, and vendors are adding controls such as IBM’s agent identity tools and previewed kill switch. Regulators are paying attention too: UK regulators say AI agents are already covered by existing law.

Why the ART result is a better guide to agentic AI than the hype

The ART study is a useful test of what agentic AI really adds. The agents did not invent new science on their own. They did the part of the job that people find slowest: combing through hundreds of thousands of sequences, rejecting most candidates and writing up the few that looked odd. Humans chose the question, ran the experiments and still do not claim to know what the system does. That division of labour is the realistic model for offices too. An agent that drafts the contract review, reconciles the invoices or proposes the shipping route saves hours, but the value depends on someone checking the output and owning the decision.

That also explains why governance has become part of the product rather than a compliance afterthought. The more steps an agent takes without asking, the more an organisation needs to know which agent acted, what it was allowed to do and how to stop it. Companies that set clear goals, limits and review points will get the productivity gains with fewer surprises; those that hand agents broad access and hope for the best are likely to learn the hard way. The next thing to watch is Anthropic’s follow-up work on what ART actually does, which will show whether agent-found leads turn into tested biology.

Update, 3 November 2026: Corrected the Anthropic figures to roughly 950 agents (not 1,000) and removed an unsourced compute-cost estimate; Anthropic has not yet published ART’s function. Removed two quotes whose speakers and affiliations could not be verified, and replaced general website links with specific primary sources.

Sources

• Anthropic, Claude discovers a novel enzyme system with CRISPR-like repeats, 23 September 2026
• Anthropic, Autonomous AI agents discover reverse transcriptases with tandem repeat arrays (technical report)
• Dario Amodei, Machines of Loving Grace
• NIST, AI Risk Management Framework
• vLLM documentation
• llama.cpp on GitHub
• LiteLLM documentation
• Unsloth
• Stanford HAI, 2026 AI Index Report (further reading)