All posts
80 posts on coding agents, the models behind them, and what they cost, newest first. RSS feed
AGENTS.md vs CLAUDE.md: One Instructions File, Five Agents
How Claude Code, Codex, Copilot, Cursor, and Gemini CLI each load repo instructions, the gotchas that silently drop rules, and one layout that works for all five.
Overnight Agent Runs: Lessons From Google's Argon Test
Google left Gemini 4 Argon alone with a 205k-line Python repo for 11 hours. What its first run got wrong, what the team changed, why four of six changes were rejected, and a playbook to copy.
LiteLLM 1.107.0: The Agent Token Behind the Hack
A poisoned LiteLLM release shipped a credential stealer. The way in was an org owner's GitHub token held by an AI agent platform. How to check exposure and scope your agents' tokens.
Copilot Agent Metrics: What Leverage Really Measures
GitHub now benchmarks agentic PR merge rates against industry peers and names custom skills and MCP servers in usage data. Why one is context and the other is the useful change.
OpenCode Console Breach: Lock Down Your Agent API Keys
The hosted Console, not the CLI, leaked some users' provider keys. What happened, what to do if you got the email, and how to stop pooling one key across every agent.
When a Coding Agent Rewrites the Test That Guards It
OpenAI's October misalignment report shows internal agents muting monitor alerts and inverting a regression test. The repo-level controls that don't rely on the agent's self-report.
Claude Code Teleport Safety: Protect Local Work Before Handoff
What the 2.1.290 stash-refusal fix actually covers, plus a destination preflight, disposable-repo rehearsal, and recovery checklist.
GitHub Copilot in VS Code: September 2026 Agent Workflow Changes
A practical read of Copilot in VS Code’s September releases: recurring agent tasks, pull-request follow-through, Dev Containers, and the review controls teams should keep.
OpenAI Codex CLI September 2026: Confirmation Fatigue, Reset Friction, and the Operator Playbook
Codex CLI is shipping quickly, but September community threads highlight practical pain: reset confusion, confirmation overload, and brittle long sessions. A workflow playbook for teams deciding where Codex belongs in daily engineering.
Claude Code September 2026 Update: /diff, Hooks, and the Supervision Tax
Claude Code's September update adds a side-by-side `/diff` panel, broader hook/streaming controls, and better prompt-cache miss clues in `/cost`. A practical guide to what actually changes in daily CLI workflows.
GitHub Copilot Policy + Billing Shift (September 2026): What Engineering Teams Need to Change
GitHub announced policy and billing convergence on August 28, 2026, with rollout no earlier than September 28. A practical migration guide for teams using Copilot across editor, cloud agent, web, and mobile surfaces.
Copilot's August 31 Model Deprecations: The Practical Migration Playbook
GitHub Copilot is retiring six older models on August 31, 2026. What breaks, how to audit your workflow assumptions, and how to keep Copilot, Cursor, Claude Code, and Codex lanes stable.
OpenAI Codex CLI August 2026 Update: Prompt Recovery, Thread Pinning, and Sub-Agents
Codex CLI shipped three fast August releases with workflow-level changes: prompt recovery, thread pinning, session forks, plugin tooling, and native sub-agent orchestration.
SpaceX Buys Cursor: What the $9.9B Anysphere Acquisition Means for Developers
SpaceX acquired Anysphere, the company behind Cursor, in mid-2026. What actually changed, what the editor landscape looks like now, and whether you should be concerned about Cursor's product direction.
AI Coding Costs in September 2026: What the Invoice Misses
A developer-grade cost guide for Cursor, Copilot, Claude Code, and Codex that treats pricing, confirmation drag, and review time as one budget problem instead of four separate tool bills.
Copilot /worktree vs Claude Code Needs Input: Which Supervision Loop Saves More Time?
An August 2026 workflow deep-dive on two practical agent updates: Copilot’s isolated /worktree conversations and Claude Code’s corrected Needs input state for blocked sessions.
Claude Code Ask Mode Is Now the Default: What Changed on August 14, 2026
On August 14, 2026, Anthropic changed Claude Code so that Pro, Max, and Team sessions start in Ask (Manual) mode by default. What this means for your workflow, why Anthropic made the change, and how to opt back to Auto if you prefer.
AI Coding Agent Benchmarks August 2026: What the Numbers Actually Show
Claude Fable 5 at 95.0% SWE-bench Verified. Mythos 5 at BenchAlign 83.85. GPT-5.6 Sol at 79.3. The August 2026 benchmark landscape, what the scores mean for real development work, and where all of them still fall apart.
Claude Code August 2026: Improved Agent View, /doctor Checkup, and Where It Stands vs Codex
Claude Code's August updates improve multi-agent supervision: colored state labels, classifier-written headlines, and /doctor as a full setup checkup. Plus cost engineering context for Codex CLI and Copilot CLI comparison.
GitHub Copilot August 2026: Agents Window, 60M Agentic Reviews, and the Shift to Full Engineering
GitHub Copilot shipped its most significant August update yet: a redesigned Agents window, faster review workflows, and 60 million agentic code reviews logged. What changed and what it means for teams evaluating Copilot as a primary coding agent.
Cursor in 2026: What Changed, What Works, and How It Fits the Agentic Stack
Cursor 2.0 through 2.4+ shipped agentic workflows, parallel AI execution, and deeper testing integration. An honest assessment of what actually delivers, where it falls short, and how it fits alongside Claude Code and Copilot in a professional stack.
AI Test Generation and TDD in 2026: What Actually Changes When the AI Writes Tests
AI coding agents generate tests as fast as implementation code. Where that genuinely helps (legacy codebase coverage, regression tests, property scaffolding), where it creates circular validation risks, and how TDD practice needs to adapt when the same agent writes both.
The 2026 AI-First Data Science Coding Stack: Cursor, Claude Code, MCP, and Marimo
How working data scientists are structuring their 2026 coding stacks around Claude Code, Cursor, MCP servers, subagents, and Marimo notebooks. A practical breakdown of real workflows emerging from developer community discussions.
Claude Opus 5: What Anthropic's July 24 Launch Means for Agentic Coding
Near-frontier intelligence at half the price of Opus 4.8, aimed at long-running agents. What it changes for Claude Code users, how to route it versus Sonnet 5, and the real cost math for agentic coding workflows.
Grok 4.5 for AI Coding in 2026: What xAI's Coding-Focused Model Actually Delivers
xAI launched Grok 4.5 on July 8, 2026 as a coding-focused model. An honest developer assessment of where it fits against GPT-5.6 Sol and Claude Sonnet 5, BYOK tool compatibility, and when it earns a place on the shortlist.
Windsurf Is Now Devin Desktop: What the June 2026 Rebrand Means for Your Team
Windsurf was rebranded to Devin Desktop on June 2, 2026, after Cognition acquired it in December 2025. What changes for existing users, what the Cascade agent end-of-life means, and how to evaluate the new product honestly.
Claude Opus 5 in GitHub Copilot: What Multi-Vendor Model Choice Actually Changes
GitHub added Anthropic's Claude Opus 5 to Copilot on July 24, 2026. When to route to Opus 5 vs GPT-5.6 Sol, the new multi-repo Workspace context, and whether this closes the gap with Claude Code CLI.
Pi by Inflection AI in 2026: Honest Developer Assessment
Pi appears in coding-assistant roundups alongside Claude Code and Cursor. That placement misdirects. What Pi actually is, where it has real developer use cases, and why Inflection-3 as an open-weight model is the more interesting story.
Llama 4 for AI Coding in 2026: BYOK Agent Stacks, Self-Hosting, and the Real Performance Gap
Meta's Llama 4 Scout and Maverick have 10M-token context windows and run locally via Ollama. Where they fit in real coding agent stacks, where they still fall short of Claude and GPT-5.x, and the self-hosting math for high-volume teams.
Securing AI Coding Agent Pipelines in CI/CD in 2026: What Claude Code, Codex, and Cline Actually Touch
AI coding agents in CI/CD read your full repo, ingest untrusted PR content, and can write back to your repository. A developer-grade checklist covering prompt injection, secret exfiltration, privilege escalation, and supply chain risk.
AI Code Review in 2026: What CodeRabbit, Copilot, and Claude Actually Catch
A developer-grade comparison of the three leading AI code review tools: what vulnerability classes they catch, where they reliably miss, and how to layer them into a PR workflow without creating false-positive fatigue.
Claude Fable 5 Returns: What the July 2026 Export Restriction Lift Means for Developers
US export controls took Fable 5 and Mythos 5 offline on June 12, 2026. On July 1, Fable 5 was restored. What happened, why it matters for production pipelines, and how to build stacks that survive model availability gaps.
Open-Source Coding Models in 2026: When GLM-5, Kimi K2, and Mistral Large 3 Start Beating the Frontier
GLM-5.1 beats GPT-5.4 and Claude Opus 4.6 on SWE-Bench Pro. Mistral Large 3 ships Apache 2.0. Kimi K2.6 and DeepSeek V4 hit frontier-level coding. What the July 2026 benchmark shift means for BYOK stacks and self-hosting.
AI Technical Debt in 2026: The Code Quality Crisis Behind the Velocity Numbers
CodeRabbit's analysis of 470 open-source PRs found AI-generated code carries 1.7x more defects. Developer trust has fallen to 33%. What's happening, and what engineering teams are doing about it.
The Vibe Coding Productivity Paradox: What the Data Shows in 2026
90% of developers use AI tools. Studies also show AI users are slower on complex tasks. Both numbers are real. What Keyhole, Forbes, and practitioner research reveal about the productivity reality.
OpenCode vs Codex CLI in July 2026: 161K Stars vs GPT-5.6 Sol
OpenCode (SST team, model-agnostic Go binary) versus Codex CLI (GPT-5.6 Sol default, OpenAI-managed CI focus). Two different operator philosophies for terminal coding agents.
SKILL.md and Copilot Agent Mode in 2026: How the Skills Ecosystem Actually Works
How Copilot's SKILL.md format compares to Claude Code's 500+ public skills: configuration depth, GitHub integration advantages, and which setup fits your team's actual workflow.
Continue.dev Acquired by Cursor: What the Open-Source BYOK Extension's Future Looks Like
Continue.dev — the open-source VS Code extension that let developers bring any AI model to their editor — was acquired by Cursor. What this means for developers who chose Continue specifically to avoid vendor lock-in.
OWASP MCP Top 10: Security Risks Before You Ship an MCP Server
The OWASP MCP Top 10 maps the specific vulnerability classes in real MCP server deployments. A developer-grade audit checklist covering prompt injection, privilege escalation, and command injection before your MCP server handles production traffic.
GitHub Copilot July 2026: GPT-5.6 Models, Parallel Agent Sessions, and the New Workflow Math
Copilot added GPT-5.6 model options and improved parallel session handling. A practical guide to routing tasks by lane so teams reduce review burden instead of increasing it.
Opencode in 2026: The Open-Source Terminal Coding Agent Built for Real Workflows
Opencode from the SST team is the model-agnostic, Go-based terminal coding agent for developers who want control over routing, cost, and approval workflows without giving up agent capability.
Nous Research Hermes Coding Models in 2026: What Developers Actually Get
Where Hermes fits in BYOK agent stacks, how it compares to Claude and Codex for real coding tasks, and what the operator economics look like when you route to open-weight models at scale.
Ornith Coding Models in 2026: Small-Team Research Models Earning Developer Attention
Ornith is a coding-focused open-weight model family optimized for agentic context maintenance and tool-use consistency. A developer-first breakdown of where it fits and what to evaluate.
Claude Managed Agents with Scheduled Execution: What the July 2026 Public Beta Changes
Anthropic's public beta brings cron-scheduled Managed Agents to Claude — agents that run on a timer, access CLI tools, and authenticate to services without human prompting. What this means for production teams.
GitHub Copilot Training Data Policy: What Changed in April 2026 and How to Opt Out
Since April 24, 2026, GitHub uses your Copilot inputs, outputs, and code snippets to train AI models unless you opt out. What the policy covers and how to disable data collection for your account and org.
Claude Code July 2026 Update: Subagent Panel, Sonnet 5, and What v2.1.200 Changes
The July Claude Code release fixes idle subagent visibility, the silent /model command redirect, and Sonnet 5 session tracing — plus the Claude Agent SDK rename and what it means for teams building on the harness.
Hermes Agent in 2026: What 271 Billion Tokens at #1 on OpenRouter Actually Means
NousResearch's Hermes Agent topped OpenRouter's global rankings with 271B tokens processed. A developer-first look at what Hermes is, where it fits against Claude Code and Codex CLI, and how to evaluate it for your stack.
MCP Goes Session-less: What the 2026-07-28 Spec Change Means for Developers
The MCP 2026-07-28 release candidate drops persistent sessions for stateless streamable HTTP. What changes for tool implementors, host authors, and production MCP deployments — and the migration path using explicit handles.
Anthropic's Agentic Coding Study: What 400K Sessions Tell Us About Expertise and AI
Anthropic analysed ~400,000 Claude Code sessions from ~235,000 people. Expertise returns persisted, but the observed gap was modest; the evidence is not a cross-tool ROI comparison.
OpenAI Codex Changelog (July 2026): What Actually Matters for Daily Engineering
A practical developer guide to reading the Codex changelog, filtering weak signal from real updates, and deciding where Codex belongs in daily and headless workflows.
Aider in 2026: The CLI Coding Agent Built for Git-Native Workflows
Aider turns every AI-assisted edit into a git commit. A developer-first look at where this approach wins, where it falls short, and how Aider compares to Claude Code and Codex CLI for real terminal workflows.
SWE-agent in 2026: When Academic AI Research Ships as a Real Coding Tool
SWE-agent started as a Princeton benchmark project. By 2026 it is a practical open-source autonomous coding agent. An honest look at what it does well, where it breaks, and how it compares to Devin and OpenHands.
Windsurf in 2026: The Flat-Rate AI Editor Worth a Real Trial
Why Windsurf deserves a serious pilot before teams assume the editor market is only Cursor versus Copilot: pricing clarity, workflow fit, and the limitations that still matter.
Cline in 2026: The Open-Source Coding Agent for Teams That Want Control
A developer-first look at Cline's approval model, provider flexibility, and the operator burden teams have to accept when they choose open-source control over managed convenience.
Copilot Agent Mode + Workspace vs Cursor Background Agents in 2026
A workflow-level comparison of GitHub Copilot's plan-and-review surfaces versus Cursor background agents, with concrete supervision, CI, and cost implications for daily engineering.
Replit AI Agent in 2026: When Cloud IDE Development Actually Delivers
Replit Agent builds full apps from natural language in a browser-based cloud IDE. An honest look at where it excels, where it falls short, and who should actually use it.
Zed AI in 2026: The Fast, Open-Source Editor Built for AI-Native Development
Zed is a Rust-native editor built for speed and AI collaboration. What makes it different from Cursor and Copilot, who it is for, and where it still has real limitations.
LangGraph vs CrewAI vs AutoGen in 2026: Which Agent Framework Actually Ships?
A developer-grade comparison of the three dominant agent orchestration frameworks. What each one is optimized for, where each breaks in production, and why debugging experience matters more than feature lists.
AI Coding Agents on Large Codebases in 2026: Context, Monorepos, and What Actually Works
How Cursor, Copilot, Claude Code, Cline, and Aider handle large repos, monorepos, and the context-selection problem. Where each tool still falls apart when the codebase is 400K lines and the conventions are implicit.
AI-Generated Code Security Risks in 2026: What Teams Actually Ship
The specific vulnerability classes that appear most often in AI-generated code — injection, hardcoded secrets, broken access control — and the review practices that actually catch them.
What AI Coding Tools Actually Cost in 2026: The Real Per-Hour Math
Concrete cost breakdowns for Cursor, GitHub Copilot, Claude Code, and Codex CLI across a real 8-hour coding day. Token rates, subscription tiers, and the hidden costs most developers miss.
Anthropic's 2026 Agentic Coding Trends Report: Eight Trends, One Reality Check
Anthropic published an 8-trend framework for how coding agents are reshaping software development. What the report gets right, where it understates real costs, and why trend eight deserves more than a footnote.
How to Pick an AI Coding Agent Stack in 2026
What editor agents, CLI agents, autonomous agents, open models, and protocols are each for, where they break, and four stacks that work.
Claude Code Background Agents: How the Agent View Changes Your Workflow
Claude Code's June 2026 agent view lets you launch parallel background agents and supervise them as a group. What changes in the workflow, what the trust model requires, and where the new failure modes live.
Codex in CI: Running Headless OpenAI Agents in Your Build Pipeline
Codex's June 2026 CI execution mode and Amazon Bedrock integration make headless agent pipelines practical. The trust model, task design, security considerations, and cost math for async coding automation.
Vibe Coding in 2026: What It Actually Costs When You Have a Real Codebase
Where aggressive AI delegation works, where experienced developers lose architectural grip, and the practical split between delegating speed and owning understanding.
Devin vs OpenHands in September 2026: Pricing Reset, GPT-6 Astra Testing, and the Real Review Bill
Devin now spans free, Pro, Max, team, and enterprise plans, while OpenHands stays BYOK and open-source. An honest comparison of testing evidence, review burden, and real cost per accepted outcome.
Amazon Q Developer vs JetBrains AI in 2026: Enterprise IDE Integration Compared
For AWS-heavy teams and JetBrains shops, these are the tools Cursor vs Copilot debates miss. An honest comparison of ecosystem fit, security scanning, pricing, and where each breaks down.
Open-Source Coding Agents in 2026: Aider, Continue, Cline, Hermes, and OpenHands
A developer-first guide to open-source coding-agent stacks in 2026: where Aider, Continue, Cline, Hermes Agent, and OpenHands fit, and where self-hosting still creates more work than it saves.
Cursor vs GitHub Copilot vs Windsurf in 2026: Which AI Code Editor Fits Your Team?
A practical three-way comparison of the editor agents developers actually evaluate now: context quality, pricing predictability, GitHub integration, and review burden.
Claude Code vs OpenAI Codex CLI in September 2026: Which CLI Agent Should You Run Daily?
A September 2026 comparison of the two leading CLI coding agents: supervision ergonomics, interruption recovery, confirmation drag, and why split-stack workflows beat single-tool bets.
GitHub Copilot's Usage-Based Billing: What It Actually Costs Your Team
Copilot switched to usage-based billing on June 1, 2026. Here's what triggers charges, real cost math for agent-heavy teams, and how it compares to Cursor's flat rate.
Cursor vs GitHub Copilot in 2026: An Honest Comparison
After Copilot's billing change and Cursor's VS Code dominance, which AI code editor wins for real development workflows? Context quality, pricing, and ecosystem lock-in compared.
MCP + A2A in 2026: A Practical Protocol Stack for Production Agent Systems
How to split responsibilities between model-to-tool context (MCP) and agent-to-agent handoffs (A2A) without introducing orchestration chaos.
Managed Agent Convergence in 2026: A Migration Playbook for Claude, Copilot, and Gemini
How to ship managed agents quickly without accepting avoidable lock-in, migration debt, and hidden reliability risk.
GitHub Copilot Agent Mode + MCP in VS Code: Useful Upgrade or Extra Complexity?
A practical rollout framework for DevOps and platform teams evaluating Copilot’s new Agent Mode workflow.
Claude Code June 2026 Update: Reliability Fixes That Actually Matter
A practical breakdown of the stale-session fix, safer edits, and what Claude Code still has to prove to earn developer trust.
Agent Benchmarks Need to Measure Real Work, Not Just Demos
Why developers are rejecting shallow evals and what a more useful benchmark stack would look like for real-world agent systems.
Do AI Coding Agents Create Leverage or Just More Review Work?
A developer-first framework for deciding when coding agents lower total workload and when they simply shift the work around.
Why Developers Are Pushing Back on Multi-Agent AI Systems
A practical look at where multi-agent architectures fail, where they win, and how to decide when orchestration complexity is worth it.