AgentsAutonomy & tool use
AI agents outnumber staff at startups like Anything as workers shift to supervising bots
Startups such as Anything now run with far more AI agents than people, and Lindy has cut staff as agents took over work. Employees increasingly supervise error-prone bots, some customers push back, and large-firm adoption is outpacing measured profit gains.
From the The Guardian · AI article of October 29, 2026 and cited sources: Agents are already standing in for staff at some startups. [1]
SOURCES[1] What is it like to work when your co-workers are AI agents? theguardian.com
AT Pulse research deskPublished · Updated · 5 min read · AI-assisted analysisAI-assisted, not human-edited. Methodology Corrections Who owns this site
- What happened
- Agents are already standing in for staff at some startups. [1]
- Original source
- The Guardian · AI article, independent perspective
- Brief by
- AT Pulse research desk AI-assisted; publisher and contact on the desk page
- When
- October 29, 2026 article published
- Covered
- Oct 30 edition
- Evidence
- 16 sources across 15 sites
- 1What is it like to work when your co-workers are AI agents?theguardian.com
- 2McKinsey: The State of AI in 2026mckinsey.com
- 3Fortune: MIT report, 95% of GenAI pilots failingfortune.com
- 4PYMNTS: Anthropic Coworkpymnts.com
- 5Aragon Research: Claude Coworkaragonresearch.com
- 6PCWorld: Microsoft's Scoutpcworld.com
Reach: not measured yet
Agents are already standing in for staff at some startups. Anything has 15 employees working alongside about 150 agents. Lindy shrank from nearly 50 people to about 20, and its agents now handle announcements, social posts and marketing videos once produced by roughly 100 people in its creator program. [1]
Human work is moving toward managing agents and checking what they produce. Recruiting agents mixed up candidates with similar names and made up résumé details. Anthropic's internal report estimates that reading all of its agents' output would take non-technical staff half of their work week. [1] [9]
Agents that deal with outside people carry reputational risk. Callers flirted with River's sales agent and pushed it for bigger discounts. Coval bars agents from talking to customers, and one contractor said recruiter bots put him off those employers. [1]
field reporting and surveys show agents taking over tasks while profit impact stays flat.
Read the full assessment
Implication: as OpenAI, Microsoft and Anthropic ship always-on workplace agents, firms should budget for human review time and treat customer-facing agents as a reputational risk.
Executive brief
At the AI startup Anything, 15 people work alongside about 150 AI agents, and its founder, Dhruv Amin, expects every company to have more AI co-workers than human ones within a year (Guardian). The Guardian's reporting finds agents becoming a status symbol in San Francisco, and they are already replacing jobs: Lindy went from almost 50 staff to about 20. The same reporting shows the costs. Employees have become managers of agents that make mistakes, and some customers and job candidates are put off when they deal with bots. Large-firm adoption is rising, according to McKinsey, but measured financial returns have not moved.
What changed and event timeline
MIT finds most GenAI pilots stall
MIT NANDA's "GenAI Divide" report says 95% of organisations saw no measurable return. It blames brittle workflows and poor integration more than model quality ().
Anthropic launches Claude Cowork
A research-preview desktop agent for non-technical knowledge workers, built on the same foundations as Claude Code (;).
Microsoft unveils Scout
An always-on "personal agent for work" announced at Build, tied into Teams and Outlook. Access starts with enterprise Frontier users ().
OpenAI ships Dots
Persistent agents announced at DevDay that keep working after a chat ends. They run on GPT-6 Astra, each with its own cloud computer and browser (;).
Anthropic reports on its own workforce
By August, interactive agent hours outnumbered employee hours six to one. Junior technical hiring slowed ().
Guardian profiles agent-heavy workplaces
Reporting from startups and a Los Angeles production company on agent "co-workers", layoffs, oversight burdens and pushback from customers ().
Capabilities and access
Read the full section
- OpenAI Dots (GPT-6 Astra): connects to 4,000+ apps and works through ChatGPT, Slack and Teams. It learns user preferences, has custom action rules and runs an auto-review of consequential actions. The first Dot comes with Pro and Business Premium plans (VentureBeat); Pro starts at $100/month (PCWorld).
- Microsoft Scout: Frontier-only for now, and the first of Copilot's "Autopilots" (PCWorld).
- Claude Cowork: a desktop research preview (PYMNTS).
- Lindy, Viktor, River: startup agents covering sales, admin, "chief of staff" work and sales demos (Guardian).
Technical analysis for researchers and developers
- Pattern: the new products share a design. No vendor evaluation results were published with these launches.
- Reproducible baseline: TheAgentCompany simulates a self-hostable company with 175 tasks across engineering, project management, HR, finance and admin, plus simulated colleagues.
- Failure modes reported in the field: recruiting agents mixed up candidates with similar names and invented parts of résumés (Guardian).
Read the full section
- Pattern: the new products share a design. Each agent is persistent, runs in its own sandboxed compute, connects to tools through plugins, and gates risky steps behind human approval. Dots, for example, uses read-only tools for background research and requires human sign-off for sensitive actions such as password changes (VentureBeat). No vendor evaluation results were published with these launches.
- Reproducible baseline: TheAgentCompany simulates a self-hostable company with 175 tasks across engineering, project management, HR, finance and admin, plus simulated colleagues. The best agent completed 24% of tasks on its own. Those results come from older models.
- Failure modes reported in the field: recruiting agents mixed up candidates with similar names and invented parts of résumés (Guardian).
Claims and evidence
- Agents taking over work. This is the company's own description (Guardian).
- About 40% of organisations with $1bn+ in revenue are scaling agents, up from 27%. The survey covered 1,719 people, May–June 2026 (McKinsey; summary).
- BCG's 10–15% workforce cut. The reviewed sources do not corroborate this figure. Other BCG coverage stresses reskilling over job cuts (Asiae).
Read the full section
- Agents taking over work. Lindy's agents now handle product announcements, social media posts and marketing videos once done by about 100 people in its creator program. This is the company's own description (Guardian).
- Large-firm adoption. About 40% of organisations with $1bn+ in revenue are scaling agents, up from 27%. Only 37% of all respondents report positive EBIT impact, unchanged from 2025. The survey covered 1,719 people, May–June 2026 (McKinsey; summary). The Guardian describes this as "used agents in some capacity", which is broader than "scaling".
- BCG's 10–15% workforce cut. The reviewed sources do not corroborate this figure. Other BCG coverage stresses reskilling over job cuts (Asiae).
- Anthropic's internal data is self-reported and comes from a single company (Shafaqna).
Context and prior work
Earlier agent benchmarks tested isolated tasks. TheAgentCompany moved closer to real workplaces. MIT NANDA found that GenAI tools improved individual productivity but rarely showed up in profit and loss (Mind the Product). Lindy relaunched in February 2026 as an autonomous executive assistant (Lindy AI review).
Limitations, safety and contested findings
- Oversight load: Anthropic's report says reading all of its agents' output would take non-technical staff half the work week.
- Questioned motives: Yale's Julián Posada says the report reads like the work of a company seeking investment.
- Off-script customers: callers flirted with River's sales agent and pushed it for bigger discounts (Guardian).
Read the full section
- Oversight load: Anthropic's report says reading all of its agents' output would take non-technical staff half the work week. Typical employees also spend 20% more time in meetings (Let's Data Science).
- Questioned motives: Yale's Julián Posada says the report reads like the work of a company seeking investment. Other critics point to the lack of independent scrutiny (Shafaqna).
- Off-script customers: callers flirted with River's sales agent and pushed it for bigger discounts (Guardian).
- Pushback on bot outreach: Coval bans agents from talking to customers, and a contractor says recruiter bots drove him away from those employers (Guardian).
Business and practitioner implications
- Hiring: the reviewed sources describe hiring priorities shifting from years of experience toward the ability to delegate, judge and orchestrate several agents at once.
- Review time: budget for human checking, which can cancel out the time agents save.
- External use: customer-facing and recruiting agents carry reputational risk. Be open about when people are dealing with an agent.
Read the full section
- Hiring: the reviewed sources describe hiring priorities shifting from years of experience toward the ability to delegate, judge and orchestrate several agents at once. Entry-level roles are under pressure (Guardian; Let's Data Science).
- Review time: budget for human checking, which can cancel out the time agents save.
- External use: customer-facing and recruiting agents carry reputational risk. Be open about when people are dealing with an agent.
- ROI: measure return at the firm level. Agent adoption has grown faster than EBIT impact (McKinsey).
The source trail.
Sources (16)
What is it like to work when your co-workers are AI agents?
Article text retrieved; extracted text may omit tables or interactive elements.
theguardian.com