Bilingual

Deep Agent Code Capabilities
3 days ago
- dcode is an open-source coding agent built on the Deep Agents SDK.
- It works with any large language model, allowing switching between providers mid-session.
- Features include persistent memory across conversations, customizable skills, and approval controls for code execution.
- Supports various built-in tools for file operations, shell execution, web search, and task planning.
- Offers interactive and non-interactive modes, with CLI commands for agent and session management.
- Includes capabilities like subagents, human-in-the-loop approval, and LangSmith tracing for observability.
- Configuration is stored in ~/.deepagents/ with per-agent directories for memory and skills.
- Provides slash commands in interactive sessions for model switching, skill invocation, and memory management.
- Supports remote sandboxes for code execution and MCP tools for external integrations.
- Non-interactive mode allows automated task execution with safeguards like turn and time limits.
Claude's Criminally Bad Electron Mac App Is an Inside Job
2 days ago
- Anthropic's Claude desktop app for MacOS is built on Electron, which is criticized for not providing a native Mac experience.
- The article suggests that despite Anthropic's advancements in AI coding tools like Claude Code, the team chose Electron due to engineers' familiarity, particularly Felix Rieseberg, a key Electron developer.
- The post argues that Electron limits app functionality and performance, citing examples like Notion migrating to SwiftUI for better performance, and questions why Claude hasn't followed suit.
$85,000 in tokens later: What I learned from scaling agentic coding at Lovable
a day ago
- The author's AI token usage increased from $600/month to $25K/month after joining Lovable, with an overall spend of $85K since January.
- Development evolved from a solo dev with few agents to a system with a dedicated agent managing swarms of subagents, boosting merged PRs from 20-30 to 150+ per week.
- Human review is now reserved for high-impact changes like RFCs and ADRs, not traditional code review, shifting focus to system design and infrastructure decisions.
- AI workflows classify PRs by risk level (high, medium, low), enforcing human review for high-risk changes and using AI reviews for others to manage scaling and risk.
- Different AI review tools were tested: CI reviews serve as a safety net, while local reviews with subagents handle bulk work, though sharing setups between people remains unsolved.
- PRs remain useful as an abstraction for unverified changes, and stacked PRs help manage context and risk, as large PRs are not reviewable by AI or humans effectively.
- Managing human and agent attention involves using fast mode for interactive tasks, ensuring task autonomy to minimize permission requests, and batching validation work per PR stack.
- Agents require external task trackers like Beads, subagents for large tasks, small PRs to avoid context limits, and new context per task for basic hygiene.
- Task tracking for humans and agents should be separate to avoid clutter from agents' ephemeral text, keeping systems like Linear clean.
- Improving the development process involves agents writing skills, reflecting on past work, sharing context via git, and encouraging cultural sharing, though sharing skills is still challenging.
- The author merged 293 PRs in a week with no production defects, viewing the productivity gains as beyond unassisted development, despite token costs, with focus on outcomes over efficiency.
- The journey is described as fun, with anticipation for future innovations, and Lovable is hiring for those interested in this approach.
Mark Zuckerberg tells staff that AI agents haven't progressed enough
5 hours ago
- Meta's AI agent development pace has not accelerated as expected, despite workforce reassignments to AI groups.
- CEO Mark Zuckerberg acknowledged that recent job cuts, involving layoffs and reassignments, were not executed cleanly.
- Zuckerberg expressed concerns about adapting quickly to tech industry changes, driving the restructuring.
- The anticipated benefits of Meta's AI-focused structure have not yet materialized, though improvements are expected within 3-6 months.
- Meta's AI unit has been described negatively by some engineers as a demanding environment.
- Meta plans significant investment in AI infrastructure, with spending potentially reaching $145 billion this year.

Hasty Briefsbeta

#ai development

Deep Agent Code Capabilities

Claude's Criminally Bad Electron Mac App Is an Inside Job

$85,000 in tokens later: What I learned from scaling agentic coding at Lovable

Mark Zuckerberg tells staff that AI agents haven't progressed enough