A new artificial intelligence assistant is being trained to digest everything a software team produces—code, emails, project reports, Slack conversations—and learn how those ingredients combine into a finished application. Its creators believe that an AI capable of internalizing a company’s full development workflow will not only become a smarter programming partner but also take a step toward machines far beyond current large language models.
Reflection, a Brooklyn startup founded by veterans of top research teams at Google, unveiled this agent, named Asimov. The firm set up shop in Williamsburg, just across the street from a trendy pickleball club, to work on the idea that coding is the most direct way for an AI to shape the digital world. By giving Asimov access to all the artifacts surrounding a software project, Reflection hopes to teach it not just to spit out snippets of code but to grasp the full arc of a feature or product launch.
Misha Laskin, Reflection’s chief executive, argues that mastering code creation and maintenance is essential for next-generation AI helpers. He points out that while many teams are busy building agents that browse the web or interact with users through traditional interfaces, large language models struggle to apply those skills in a complex engineering context. Laskin’s résumé includes work on Gemini and agent research at Google DeepMind, but his current focus is training AI to see how lines of code, meeting notes, bug reports, and design documents all come together.
Laskin says Asimov has been set up to spend far more time studying existing code than churning out new functions. “Everyone is really focusing on code generation,” he told me. “But how to make agents useful in a team setting is really not solved. We are in kind of this semiautonomous phase where agents are just starting to work.” By emphasizing analysis over output, Reflection hopes the system will better understand the context for any request and suggest changes that fit a project’s requirements and style guidelines.
Under the hood, Asimov isn’t a single monolithic model but a collection of collaborating agents cloaked inside a unified interface. A set of smaller modules retrieves relevant files, snippets of conversation, and documentation, then passes these pieces to a central reasoning component. That reasoning agent weaves together the retrieved material into an answer that addresses a developer’s query, whether it’s diagnosing a failing test case or explaining how a legacy subsystem is supposed to operate.
Reflection commissioned an in-house survey comparing Asimov with Anthropic’s Claude Code running Sonnet 4. Engineers working on several large open source projects were presented with code-related challenges and asked which assistant provided the clearest, most actionable guidance. In 82 percent of cases respondents favored Asimov’s explanations, while Claude Code’s responses were preferred 63 percent of the time. Those results, Reflection says, suggest its approach can edge out some of the most advanced coding tools on the market.
Daniel Jackson, a computer scientist at the Massachusetts Institute of Technology, finds the startup’s scope of input intriguing. Since Asimov pulls in private notes, design diagrams, and casual chat along with raw code, it could capture nuances that other agents miss. He questions whether the added depth will justify the extra computing power and whether the practice of ingesting sensitive messages might raise fresh security concerns. “It would be reading all these private messages,” he says.
Reflection maintains that its multiagent design helps control compute expenses by spinning up lightweight retrieval workers only when needed. The company also says it uses a locked-down environment that isolates corporate data more securely than typical cloud-based software services, though it declines to share full architectural details for security reasons.
At Reflection’s Brooklyn offices I spoke with Ioannis Antonoglou, the chief technology officer whose past work on reinforcement learning powered landmark systems like AlphaGo. At Google DeepMind he pioneered techniques that let AI learn through trial and error, guided by rewards and penalties. Today, Antonoglou applies those same ideas not to board games but to code construction, letting Asimov refine its reasoning abilities over thousands of simulated development cycles.
Reinforcement learning from human feedback has become a standard way to improve large language models. By rating generated text and steering the model toward more accurate or polite responses, trainers can shape an LLM’s behavior. Reflection takes this a step further, using reinforcement signals to teach Asimov how to break down complex software problems into smaller steps, solve each part, and stitch together a reliable solution. Although Asimov currently uses open-source foundation models, Reflection is rolling out custom models fine-tuned with its own reinforcement learning pipelines. These models learn to assemble a working application rather than win a game of Go.
To feed that training process, Reflection draws on two sources. It employs human annotators to craft realistic software scenarios and also generates synthetic examples by having its agents simulate interactions. The startup’s leaders emphasize that no customer code or private communications ever enter their external training corpus, keeping proprietary material safely behind a company firewall.
Big AI players have embraced similar tactics. OpenAI’s Deep Research tool, for example, combines expert human feedback with automated searches across the internet to produce detailed reports on scientific or technical topics. “We’ve actually built something like Deep Research but for your engineering systems,” Antonoglou says, noting that code repositories alone don’t capture the informal discussions that hold vital context. “We’ve seen that in big engineering teams, a lot of the knowledge is actually stored outside of the codebase.”
Sequoia partner Stephanie Zhan, one of Reflection’s early investors, believes the startup is operating on the same level as the most ambitious corporate labs. She praises the firm for taking a broader view of software development and for demonstrating a clear path to useful products. “It punches at the same level as the frontier labs,” she says.
With deep-pocketed giants such as Meta stepping up their own superintelligence efforts—Meta’s newly minted Superintelligence Lab has committed substantial funding—startups like Reflection face fierce competition. Yet by narrowing their focus on the nuts and bolts of coding, these smaller players hope to carve out a defensible niche.
Reflection’s leadership sketches out a future in which intelligent agents evolve into organizational oracles, able to capture every lesson a team learns over years of projects and serve as a go-to adviser. The vision extends to software that fixes its own bugs, designs novel algorithms, and even proposes new hardware optimizations without human prompting. For now, the company fields inquiries from customers who want to equip sales engineers or technical support staff with an agent that understands their own workflows. “We’ve actually been talking to customers who’ve started asking, can our technical sales staff, or our technical support team use this?” Laskin says.