Breaking

Former OpenAI Researcher Diogo Almeida Challenges AI Industry's Virtual-Employee Push

October 8, 2026 at 16:00 EDT

AI Agents
Software Dev & Coding
Foundation Models

Diogo Almeida, a former OpenAI researcher who now runs TypeSafe AI, has sharpened the AI industry's debate over virtual employees by arguing that today's large language models are optimized for assistance, not unattended work. The challenge lands as coding-agent companies sell a future of AI systems that own outcomes, while developers still describe a stubborn human-review gap.

Oct 2026 · Agentic AI

Virtual employees hit the 80/20 gap

AI companies are pitching agents that own outcomes. The current debate is whether today's models can finish work unattended, or still need people to close the last mile.

80%

Model share in a semi-automated Claude Code workflow

$48B

Cognition valuation after its September funding round

719

OpenAI math manuscripts in the public catalog

70-500ms

TypeSafe's stated Jev response-time range

The semi-automation split

80%

Model work

20%

Human finish

The hard part is not drafting most of a task. It is closing the final piece reliably enough for unattended execution.

Where autonomy stalls

Task

Outcome requested

→

Model output

Draft, decision, or code

→

Human review

The recurring last mile

→

Trusted action

Only when checks pass

Why builders are optimistic

Coding agents are drawing large funding rounds.
Typed decision models can be fast and cheap for bounded tasks.
Verifiable domains give AI clearer correctness checks.

Why skeptics are cautious

RLHF chat models are optimized for a person in the loop.
AI-written code still needs careful review.
Agents also face access barriers in real websites.

The test is unattended work

The next race is not just smarter models. It is proving that agents can act inside real workflows with checks, thresholds, permissions, and escalation paths that make hidden supervision unnecessary.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.