Devin Now Commits 89% of Cognition's Code as Company Hits $1B Annualized Revenue

AI coding agents are evolving from "assistants" into "lead programmers." According to AI Magazine, Cognition's Devin agent now commits 89% of the company's own code, while Cognition's annualized revenue has jumped from $492 million in May to $1 billion, with a valuation of $48 billion. Behind these numbers lies a quiet revolution in how software is produced.

Devin's Role: From Assistant to Lead

Cognition CEO Scott Wu says Devin's current capabilities fall somewhere between a junior and mid-level engineer. Its workflow is: plan, write, test, and commit code, with humans handling final review. This "human reviews, AI executes" model is becoming the standard paradigm for AI coding agents.

More noteworthy is enterprise adoption. Goldman Sachs, Mercedes-Benz, GE Aerospace, and NewBank are all on the customer list, with distribution partnerships covering Infosys, Cognizant, and AWS. Mercedes-Benz handed over 200,000+ lines of COBOL code to Devin, and NewBank achieved 8 to 12x efficiency gains in a large-scale migration. These cases show that Devin's value lies not just in writing new code, but in paying down technical debt in legacy systems.

Cognition acquired Windsurf in July 2025, gaining an agentic IDE that runs parallel to Devin. This completes the product matrix: Devin handles end-to-end tasks, while Windsurf integrates into developers' daily workflows.

AI Coding Agent refers to a class of intelligent agents capable of autonomously planning, writing, testing, and committing code — fundamentally different from traditional code completion tools (like early versions of GitHub Copilot). Traditional tools only offer snippet suggestions as developers type, whereas agentic coding tools can accept a high-level task description, break it down into steps, invoke terminals and editors, run tests, and iteratively fix issues until the task is complete. Devin is a representative product of this paradigm, with its "plan—write—test—commit" workflow simulating the full development loop of a human engineer. This capability depends on LLMs' code understanding and tool-calling abilities, and it also introduces new risks: agents may introduce complex changes that are hard to review, making the human final-review step essential. When enterprises adopt agents, they need to establish code review, access control, and rollback mechanisms to balance efficiency and safety.

Exa 3.0: A Search API Optimized for Agents

Exa has released Exa 3.0, an upgrade to its web search API for AI agents, now available for free on the web. The new version includes a larger embedding model and a cheaper light search type priced at $4 per 1,000 requests.

The Dynamic Highlights feature is now generally available, selecting tokens from search results rather than returning full-page text. Exa claims this uses 95% fewer tokens than full-page text while also improving answer quality. Exa runs its own crawler and independent index covering over 2 trillion URLs, but offers no self-hosting option. In eight benchmarks run by Exa itself, the new version leads Perplexity, Brave, and Parallel across the board, with the largest gap on BrowseComp at 91.8%.

Mercury 3: The Speed and Limits of Diffusion Language Models

Inception's Mercury 3 uses a diffusion architecture, filling multiple tokens at once rather than generating token by token. According to MarkTechPost, it reaches 1,068 tokens per second on Nvidia GPUs and has a 1 million token context window.

In Inception's own test table, Mercury 3 leads GPT-6 Luna and Claude Haiku 5.5 (medium setting) on 8 of 12 tests, with the biggest advantages in instruction following and agentic coding. Compared to Mercury 1, the model is over 40x larger, with average quality on standard benchmarks improving roughly 3x while speed remains above 1,000 tokens per second. However, it is only available through Inception's Nvidia GPU API — no open weights and no quantized versions.

Diffusion Language Model borrows the diffusion concept from image generation: traditional autoregressive models generate text token by token from left to right, whereas diffusion models start from noise and refine tokens at multiple positions simultaneously through multiple denoising steps. This parallel filling approach can significantly boost generation speed, since a single forward pass can update multiple tokens rather than waiting serially for each one. Mercury 3 leverages this to achieve throughput of over 1,000 tokens per second while supporting ultra-long contexts. However, diffusion language models still face challenges in training stability, long-range coherence, and fine-grained control, and typically require specially optimized inference infrastructure. Mercury 3's API-only availability with no open weights also reflects the closed nature of high-performance diffusion models in deployment today. For enterprises needing local deployment or fine-tuning, this remains a limitation.

Pragmatic Choices for Local Deployment

In the local Llama community, a user plans to buy a used HP Z640 for €550 (dual E5-2680 v4 CPUs, 64GB ECC memory), plus two RTX 3060 12GB GPUs, for Qwen 3.8 Flash, llama.cpp, PyTorch computer vision training, reinforcement learning, LoRA fine-tuning, and video generation. Open questions include the dual-slot PCIe layout, power-limit adapters for the 3060s, whether the bundled Quadro M6000 24GB is still useful under CUDA 12, and case cooling given memory tests hitting 86°C.

The same community also saw the launch of ModelFit, a real-time Mac configurator that models fit based on actual llama-bench runs, showing which large models fit in memory and the expected inference speed on each Apple Silicon configuration.

Another case: a user runs Gemma 4 31B-Q5KM on a single RTX 3060 12GB at about 2.3 tokens/sec and hopes to reach 15+ tokens/sec at 16K context. Community options include an RTX 5060 Ti 16GB on a PCIe 3.0 x1 riser with layer splitting, or a cheap X79/X99 dual Tesla P100 setup.

Agent Governance: The Enterprise's New Blind Spot

Enterprises must not only defend against external malicious agents but also track and secure their own proliferating AI agents. According to CIO, most organizations still lack basic visibility. Anthropic says 82% of organizations have discovered unregistered agents on their networks, and 57% already run agents in multi-stage workflows.

The problem is that creating an agent takes just one API call or a single "vibe coding" session by any employee, with vendor, platform-native, and in-house agents piling up in layers. Security teams need a single record for each agent covering onboarding owner, access permissions, behavior, and cross-platform capabilities. An over-privileged email agent with 50 tools attached is a typical example worth investigating. Switch from Sandbox AQ now centralizes AI agent spend, performance metrics, and security posture in a single view, giving security and finance teams unified cross-platform telemetry.

Cybersecurity experts told Think Digital Partners that autonomous agents require new digital identity frameworks, because human-centric controls fail when agents act autonomously across systems.

Unregistered Agent refers to an AI agent running in an enterprise network or cloud environment that has not been formally registered and managed by the security team. As the barrier to creating an agent drops to a single API call or an employee's casual "vibe coding," business units, platform teams, and third-party vendors may each deploy agents independently, leaving security teams without a unified inventory. These agents often have excessive access permissions (such as reading email or calling internal APIs), and their behavior is hard to audit. The key to governance is establishing a single record for each agent covering owner, permission scope, behavior logs, and cross-platform capabilities, with regular reviews. A lack of visibility means enterprises cannot promptly detect anomalous behavior, data leaks, or permission abuse. Unified telemetry platforms like Switch aim to centralize agent spend, performance, and security posture, helping security and finance teams manage them collaboratively.

llms.txt: A Sitemap for Agents

Web pages wrap useful information in navigation, scripts, and ads, and agents must spend tokens to strip it away. llms.txt is a curated markdown file placed at the site root that gathers expert-level summaries and links to key pages, serving as the map agents read first.

Context windows are still limited, and every token spent parsing web page clutter is a token that never reaches the answer. A clean markdown map lets agents judge relevance before crawling full pages. The V2 revision adds standard link relations so agents don't have to guess URLs — specifically rel="alternate" type="text/markdown" and rel="describedby". OpenAI, Anthropic, and Gemini already publish developer docs this way, and documentation platforms generate the file automatically.

The engineering trade-off is between maintenance overhead and retrieval quality. Sites that keep the file updated help agents find the right information faster; outdated files are just another stale asset at the root. If you already maintain high-quality markdown docs and want coding agents or retrieval systems to reliably consume them, publish one.

llms.txt is a Markdown file placed at a website's root directory, designed to provide LLMs and AI agents with a curated summary and key links to site content. It is similar to robots.txt for crawlers, but its purpose is not to restrict crawling — rather, to guide agents in efficiently locating high-quality content. The file typically includes a site overview, links to core pages, and brief descriptions, helping agents judge relevance before crawling full pages, thereby reducing token consumption and parsing noise. The V2 revision introduced standard link relations rel="alternate" type="text/markdown" and rel="describedby", enabling agents to directly discover Markdown versions and description files without guessing URLs. OpenAI, Anthropic, and Gemini have adopted this approach for publishing developer docs, and some documentation platforms can generate it automatically. Its value depends on maintenance quality: continuous updates improve agent retrieval efficiency, while outdated files are as good as nonexistent.

Other Notable Developments

Centallion's AI agent cut trade finance amendment time by 80% (PYMNTS). Cognizant and Google Cloud expanded their partnership to scale enterprise AI agents (HPC Wire). A security firm found that an AI agent with China ties was behind cyberattacks against Japan (The Straits Times).

IT Brief Australia notes that AI agents are changing managers' work, but accountability mechanisms haven't caught up. News Medical reported on a new AI model that can detect hidden diseases in routine CT scans. Documents obtained by Gizmodo reveal a military program using agentic AI to hack into U.S. adversary systems.

Microsoft Foundry is adding scheduled agents and smarter tool use, along with GPT-6 model integration.