Reflection AI, which wants to achieve superintelligence through autonomous coding agents, emerges from stealth with a $105M Series A and $25M in seed funding

Reflection AI is building coding agents that can function autonomously. — Two top researchers from Google's artificial intelligence lab DeepMind …

Bloomberg 2025-03-07 Shirin Ghaffary

Context & Ripple Effects

This is the starting financing and public-launch point for Reflection AI, founded by former Google DeepMind researchers and focused on autonomous coding agents. Later coverage shows the company turning that research focus into Asimov, an agent designed to work across company codebases and documentation.

The arc then broadens beyond a single agent product: Reflection AI was reported to be pursuing open-source models positioned against leading closed-source systems. That makes the initial $130M disclosed here meaningful as early backing for both an agent-led product strategy and a wider model ambition.

First-order effects

Reflection AI gains $105M in Series A capital alongside $25M in seed funding, giving its newly public team resources to develop and commercialize autonomous coding agents.
The stealth exit makes Reflection AI a visible contender for engineering teams and AI talent, while associating its approach with its founders' DeepMind background.

Second-order effects

Other coding-agent providers must contend with a newly funded entrant whose stated goal is autonomy rather than narrower code assistance, increasing pressure to demonstrate useful end-to-end engineering workflows.
The funding supports a path from agent tooling toward proprietary or open models, a direction later reflected in its reported effort to build open-source models.

Third-order effects

If agent companies increasingly control both the software-engineering workflow and the underlying models, competition may shift from standalone coding features toward integrated agent-and-model stacks.
The trajectory suggests that autonomous coding agents can become a beachhead for broader AI labs; whether that produces durable differentiation depends on real-world reliability and adoption, not financing alone.

The trend: This is one data point in the move from code copilots toward embedded, increasingly autonomous agents backed by companies seeking control of the underlying model layer.

Discussion

@mishalaskin Misha Laskin on x
Today I'm launching @reflection_ai with my friend and co-founder @real_ioannis. Our team pioneered major advances in RL and LLMs, including AlphaGo and Gemini. At Reflection, we're building superintelligent autonomous systems. Starting with autonomous coding. [image]
@gergelyorosz Gergely Orosz on x
It is somewhat amusing that most of these startups realize that coding agents are the single best area to start... as LLMs are uniquely suited to do coding well (thanks to limited grammar, lots of training data and proven success) They then usually launch something that is as
@s_oheigeartaigh @s_oheigeartaigh on x
Another one. This one claims to be aiming to build superintelligence. But nothing in the article or on their website seems to require or necessarily imply the sort of superintelligence Sutskever / Altman / Hassabis et al have in mind. Could of course mean a narrower / scoped 1/3
@gergelyorosz Gergely Orosz on x
Raising funds build a “superintelligent” or “superhuman” agent/system/AI is a sure way to raise hundreds of millions of dollars from investors for folks with the right background. The companies that complete this kind of fundraising already won: Investors are taking all risk.
@trace_cohen Trace Cohen on x
Google DeepMind has 5400 employees with 40%+ in engineering/research. Any 2 or more can leave and raise $20M+ Day 1 no questions asked. We're just getting started. [image]

Chronicles