Bilingual

Coding Agent VMs on NixOS with microvm.nix
a day ago
- The author describes creating ephemeral VMs with microvm.nix on NixOS to run coding agents like Claude Code safely, isolating them from personal files so the VM can be discarded if compromised.
- Step-by-step setup includes: configuring a network bridge (microbr) with NAT, adding microvm as a Nix flake input, declaring VMs in microvm.nix, using a base configuration for shared directories, SSH keys, an 8GB disk overlay, and cloud-hypervisor with 8 vCPUs and 4GB RAM.
- The author shows how Claude Code can automate VM creation using a skill, and notes that this setup's maintenance is minimal as it updates with the host NixOS config.
What happens when coding agents stop feeling like dialup?
a day ago
- Coding agents initially felt magical but now feel like dialup due to reliability issues and infrastructure strain.
- Token usage is exploding (50x increase on OpenRouter) because agentic workflows consume ~1000x more tokens than chats.
- Current LLM speed (30-60 tok/s) is frustrating, while faster models (2000 tok/s) could enable more unsupervised parallel approaches.
- The future of coding agents may involve parallel attempts and automated evaluation, enabled by higher tok/s.
- Infrastructure scaling faces challenges from slower semiconductor improvements, leading to potential pricing model innovations like off-peak plans.
- Developers should stay curious and adapt; experienced developers often dismiss coding agents but benefit most from them.
Why sandboxing coding agents is harder than you think
a day ago
- Sandboxing coding agents is challenging because seemingly safe commands (e.g., go test, git) can be exploited to execute arbitrary code, such as deleting user home directories through test files or using git hooks.
- Docker-based sandboxing has limitations, like accidental superuser database permissions (e.g., COPY PROGRAM) or Docker socket access enabling host-level container escapes, making throwaway VMs a more secure alternative.
- Agents' logs pose a security risk by inadvertently capturing secrets (e.g., env variables, stack traces), potentially stored on provider servers, highlighting the need for auto-scrubbing and encryption.
- LLMs enable large-scale vulnerability hunting in niche open-source projects, flipping the cost asymmetry of finding vs. exploiting vulnerabilities and increasing risks for low-attention software.
- Coding agents represent a new software execution category that blurs the line between malicious and trusted code, requiring OS-level rethinking akin to smartphone permission models.
What's going on with Gemini?
a day ago
- Google's Gemini models trail behind Anthropic and OpenAI in frontier intelligence, with Gemini 3.1 Pro falling short of leading models like GPT-5.5 and Opus 4.8.
- Gemini 3.5 Flash offers impressive speed (206 t/s) but has a 3x price increase and underwhelming coding benchmarks, making its market position unclear.
- Google's custom TPU hardware and internal model-hardware collaboration provide structural advantages for inference efficiency.
- Google lacks a coherent coding agent strategy, with multiple overlapping tools (Antigravity, Jules, Gemini Code Assist, Gemini CLI, AI Studio) and weak developer adoption compared to Anthropic's Claude Code and OpenAI's Codex.
- Google's bespoke internal software workflows may hinder its ability to develop effective agentic tools for the broader industry.
My AI Adoption Journey
a day ago
- Adopting a new tool involves three phases: inefficiency, adequacy, and life-altering discovery.
- Chatbots are limited for coding; use agents that can read files, execute programs, and make HTTP requests.
- Reproduce manual work with agents to build expertise, breaking tasks into clear, actionable steps.
- End-of-day agents can perform research, triage issues, and explore ideas while you are unavailable.
- Outsource high-confidence tasks to agents while focusing on deep work, turning off notifications to avoid context switching.
- Engineer tools and prompts to prevent agent mistakes, improving reliability through 'harness engineering'.
- Aim to always have an agent running for delegated tasks, balancing it with manual work.
Reverse-engineering is cheap now
a day ago
- Coding agents are being used to reverse-engineer and automate home devices, illustrating the impact of reduced code writing costs.
- Before coding agents, reverse-engineering home devices was possible but had poor ROI due to effort and potential future maintenance.
- Coding agents lower the cost of trying and failing, making maintenance or starting over less psychologically burdensome.
Tokenflation: When "Hi" triggers 33 tool calls
3 days ago
- Tokenflation describes the phenomenon of coding agents using excessive tokens (context, reasoning, tool calls) for simple tasks like responding to 'Hi', which can trigger up to 33 tool calls and significant waiting time.
- The real cost of tokenflation is not just API token fees (often < $0.01) but the waiting time, valued at about $0.016 per second based on a $120k salary, making total costs like $0.84 for a 'Hi' response on models like Sonnet.
- In a benchmark of 14 models, ambiguous prompts ('WTF' or 'Hi') caused more tool calls, timeouts, and failures than clear tasks ('commit'), with waiting time dwarfing API costs (e.g., 99% of bill for some models).
- Labs focus on leaderboard metrics like SWE-bench, ignoring efficiency on everyday tasks, and vendors' per-token billing creates a perverse incentive for more exploration, leading to tokenflation in real-world use.
Show HN: Uploads.sh – the missing upload command for coding agents (open-source)
5 days ago
- GitHub lacks a way for agents to show work; capture screenshots as you work, which are staged for the branch, and when a PR opens, it's furnished with a comment.
- Uploads can carry custom tags using --meta; use 'uploads find' to retrieve them by tag, organizing visuals by app part or form factor.
- Uploaded files on a branch stage automatically; 'uploads staged' shows queued items and re-running updates the same comment without breaking embedded URLs.
- Features include device frames (e.g., phone, browser, iPhone), one-shot screenshots of URLs, and shareable galleries that can span multiple PRs.
- To set up, paste the prompt into a coding agent or read the full guide.
How Do We Stop Vibe Coding?
5 days ago
- Coding agents rapidly gained popularity among software developers after Claude Code's surge in late 2025, with figures like Andrej Karpathy noting a shift away from manual coding since December 2025.
- Grady Booch views this as the third golden age of software engineering, where AI agents abstract code itself, similar to how compilers abstracted the machine.
- The core problem with 'vibe coding' is the loss of understanding, awareness of dead code, and the ability to reason about or explain architectures, leading to unreliable outputs.
- Current solutions like markdown specs, skills, Spec Kit, OpenSpec, Kiro, and test-driven development fail to address trust, offering only prompt engineering without deterministic enforcement.
- Promising approaches like CodeSpeak and Scryer focus on trust by compiling specs into code or creating a shared model between developer and agent for transparency and drift detection.
- The future likely involves moving beyond manual coding for routine tasks, with agents guided by structured, auditable intent rather than vague prompts.
Reverse-engineering is cheap now
9 days ago
- Coding agents are being used by people to reverse-engineer and automate home devices, illustrating the impact of reduced code writing costs.
- Previously, reverse-engineering home devices was possible but rarely worth the effort due to low ROI and the burden of maintaining unstable APIs.
- Coding agents drastically lower the effort and cost of trying and failing, and reduce the psychological baggage of future maintenance.
- Recent articles include: Kimi K3 and the pelican benchmark (July 16), new GPT-5.6 family (Luna, Terra, Sol) (July 9), and sqlite-utils 4.0 with schema migrations (July 7).
Understanding the Cost of Coding Agents
14 days ago
- AI cost in tokens and dollars is a major concern, with examples like Uber exceeding its budget quickly.
- Coding agents involve different token types: input, cached input, output, and reasoning tokens, each with different prices and implications.
- Cached input tokens are the cheapest but can dominate costs due to repeated use in agentic loops, while output tokens are the most expensive.
- Providers like Anthropic and OpenAI handle caching differently, affecting billing and cache optimization strategies.
- Token usage patterns vary by application; coding agents typically have high cached token volumes, unlike chatbots or RAG systems.
- AI subscriptions lack transparent token allowances, making telemetry crucial for monitoring and managing token consumption.
- Investing in good telemetry is essential for understanding costs, especially when switching between multiple AI providers.
- AI budgets depend on token volume, not just pricing; cached tokens can become a major cost due to high usage volume.
- Coding agents significantly influence token volume; choosing and configuring them carefully is as important as selecting an LLM provider.
- Forecasting future token usage requires telemetry and will be explored in a follow-up article to avoid budget overruns.
Coding agents think ahead of time
15 days ago
- Language models used in coding agents internally represent program properties like correctness and test outcomes.
- Probes on hidden states can decode program states and even predict future edit outcomes up to 25 steps ahead.
- The latent programming horizon concept suggests models have internal representations that anticipate future code changes.
- Probes transfer across benchmarks without retraining, showing external validity and encouraging interpretability research.
Compressor V2: three compression layers for a 50% LLM agent cost cut
23 days ago
- Token compression is critical for coding agents due to economic pressure and operational bottlenecks like dollar cost, latency, context window limits, and throughput.
- Edgee's Compressor V2 includes three orthogonal strategies: brevity (reducing output tokens), tool surface reduction (TSR) for MCP tool catalogs, and tool result trimming for verbose tool outputs.
- Statistical analysis using paired sign tests, bootstrap confidence intervals, and coefficient of variation shows significant savings: brevity achieves ~30% cost reduction on SWE-bench Lite, TSR reduces tokens by ~33%, and tool result trimming adds incremental 5–10% savings.
- Brevity works by eliminating meta-text in outputs, preserving cache amortization, and targeting the most expensive token class without prefix changes.
- TSR rewrites tool catalogs into a single virtual tool, reducing prefix bloat and token volume, especially beneficial for tool-heavy workloads.
- The three strategies are composable and target different layers (output, prefix, history), allowing users to enable combinations per API key based on workload shape.
Does Code Cleanliness Affect Coding Agents?
23 days ago
- The study investigates whether the cleanliness (structural and stylistic quality) of code affects the performance of autonomous coding agents.
- It introduces a minimal-pair evaluation protocol, comparing repositories identical in architecture, dependencies, and behavior but differing in static-analysis violations and cognitive complexity.
- Results from 660 trials with Claude Code show that code cleanliness does not significantly change task pass rates.
- However, agents working on cleaner code use 7–8% fewer tokens and reduce file revisitations by 34%, indicating improved efficiency and reduced computational cost.
- The findings highlight that traditional maintainability principles remain relevant in AI-driven development, influencing agent behavior alongside model choice, harness, and prompting.

coding agents

Coding Agent VMs on NixOS with microvm.nix

What happens when coding agents stop feeling like dialup?

Why sandboxing coding agents is harder than you think

What's going on with Gemini?

My AI Adoption Journey

Reverse-engineering is cheap now

Tokenflation: When "Hi" triggers 33 tool calls

Show HN: Uploads.sh – the missing upload command for coding agents (open-source)

How Do We Stop Vibe Coding?

Reverse-engineering is cheap now

Understanding the Cost of Coding Agents

Coding agents think ahead of time

Compressor V2: three compression layers for a 50% LLM agent cost cut

Does Code Cleanliness Affect Coding Agents?