AI Coding Assistants

Today's Stories

1. 1

   Haleon Partners with Microsoft on Five-Year AI and Cloud Initiative

   Haleon announced on June 1 a five-year collaboration with Microsoft to integrate Microsoft's cloud and AI tools into its consumer health operations. The partnership will deploy advanced agentic AI (self-directing AI that completes tasks autonomously), security, and identity features to scale Haleon's AI infrastructure while streamlining operations from supply chain to commercial execution. Haleon aims to reach one billion more consumers by 2030, and the partnership is designed to enable faster, more data-driven business decisions. By co-creating high-impact AI use cases with Microsoft, the company intends to enhance innovation, scientific research, and consumer insights across its global operations.

   The initiative builds on Haleon's existing use of Microsoft 365 Copilot (a productivity tool that automates routine tasks) and expands that foundation with enterprise-scale AI infrastructure. The five-year timeframe suggests a sustained commitment to transforming the company's digital operations.

2. 2

   Nvidia Wins Perplexity's CPU Choice, Expanding AI Chip Dominance Beyond GPUs

   Perplexity chose Nvidia's new Vera CPUs over x86 server chips for its multi-agent AI coding stack, delivering 1.5 times faster performance than standard server processors. Nvidia's Q1 FY27 revenue hit $81.61B (+85.2% YoY), with Data Center at $75.25B (+92%) and Networking at $14.8B (+199%). AMD's Q1 FY26 reached $10.25B (+37.9% YoY), with Data Center revenue of $5.78B (+57%). Nvidia is bundling GPUs, CPUs, networking, and software as one integrated stack, while AMD is still assembling a competing answer. The Perplexity decision signals that inference buyers default to Nvidia's Blackwell plus NVLink plus CUDA when latency matters—marking a potential moat that extends Nvidia's reach into CPU territory that x86 vendors like AMD have long defended. AMD's 184x trailing P/E leaves no room for execution stumbles.

   Whether Vera CPU adoption spreads beyond Perplexity into hyperscalers, and whether AMD's MI450 shipments in H2 2026 meet the 'exceeding expectations' language leadership used. Nvidia guided Q2 revenue to $91.0B, and analysts carry a target of $301.62.

3. 3

   IBM Bob v2 coding agent speeds up with parallel tool calls

   IBM released version 2 of IBM Bob, its AI coding agent (available as an IDE extension and shell command), with a ground-up rebuild. The notable change is that tool calls now run in parallel instead of one at a time, making the agent noticeably faster. IBM Bob competes in the same space as Cursor, Claude Code, and Copilot. A developer who tested it against their usual daily drivers found it capable of ordinary app-building tasks—in this case, building a POC that pulls articles from IBM RSS feeds and a YouTube channel to auto-summarize them into a daily digest. Faster execution makes it more viable as a replacement for workflows already handled by established agents.

   The shift from sequential to parallel tool execution is a meaningful efficiency gain, though the real test will be whether developers adopt it as their primary agent over incumbents. No pricing, availability window, or region restrictions are mentioned in the release details.

4. 4

   Agent builders need dedicated databases to track what their AI actually does

   A developer who built six AI agent systems over six months found that each one required a database to log and manage the agent's execution — recording task dispatch, evaluation runs, and results — separate from standard application databases. Tracking an agent's decisions and actions in a dedicated log lets builders evaluate performance, debug failures, and let the agent reflect on its own work. Without this visibility, understanding why an agent succeeded or failed becomes much harder.

   The practice draws on best practices from companies like OpenAI, Anthropic, Stripe, and others, suggesting that agent-specific logging is becoming a recognized pattern in production AI systems.

5. 5

   One developer stops playing messenger between AI agents, wants them to talk directly

   A developer realized that manually shuttling work between multiple AI models—pasting code plans from one agent to another for review, then arbitrating their conflicting suggestions—was turning them into a referee rather than saving time. They decided they wanted the agents to discuss the work directly instead of going through a human intermediary. When using multiple AI agents for tasks like code review or planning, the bottleneck is not moving context between them but the human overhead of collecting and reconciling their feedback. Direct agent-to-agent communication could eliminate that friction, making multi-agent workflows genuinely faster rather than just multiplying the number of prompts someone has to manage.

   The developer's core insight is that agent-to-agent discussion—having one model challenge another's suggestions in real time—could replace the manual back-and-forth that currently demands human judgment. Whether this actually works depends on whether agents can productively debate trade-offs without needing human arbitration at every step.

6. 6

   Anthropic adds browser to Claude Code, letting AI read and click external websites

   Anthropic integrated a browser window into Claude Code that allows Claude to open, read, click, and type on web pages directly within the app. The browser works like a tab-based browser, opens with a keyboard shortcut, and includes safety checks—classifiers screen write actions on external sites, and Claude will not buy anything, create accounts, or bypass CAPTCHAs without user consent. Claude can now access documentation sites, issue trackers, and other web resources during coding work without leaving the app, potentially speeding up development workflows. Organizations can restrict access through an allowlist or disable the feature entirely, giving teams control over which sites Claude reaches.

   Users who need Claude to act within their own logged-in sessions should use the Chrome extension instead of the built-in browser, which runs on a clean profile with no saved logins.