frontpage.
newsnewestaskshowjobs

Open Source @Github

fp.

Open in hackernews

My 600 Hours with AI Coding Assistants: A Practical Comparison

2•bv_dev•1y ago
After spending over 600 hours using various AI coding assistants over the past 3 months, I wanted to share my experience for those navigating this rapidly evolving landscape.

What I Mean by "Agentic Mode" First, to clarify: by "agentic mode," I'm referring to the assistant's ability to understand project context, reason through multi-step problems, and autonomously make coherent code changes across files without constant hand-holding. True agency means the tool can maintain context across interactions and execute on high-level directions.

The Current Landscape (May 2025) Augment Code - Current go-to tool despite higher costs

Strengths: Maintains context remarkably well across complex refactors; actually understands project structure; can implement feature requests that span multiple files Weaknesses: More expensive than alternatives ($30/month vs $20 for others); occasional hallucinations when venturing outside codebase context Best for: Complex refactoring tasks and implementing features that span multiple files

Windsurf - Slightly edges out Cursor for agentic capabilities

Strengths: Better context retention than Cursor; decent file traversal; good understanding of code relationships Weaknesses: Can get quite stuck in their full agentic mode as it starts editing things. While they have removed their flow credits part, it is still painful to watch it go completely out of context. Best for: Mid-size projects where you need moderate autonomy

Cursor - Popular but underwhelming for true agentic work

Strengths: Good IDE integration; clean interface; works reasonably well for single-file tasks. I like the ability to Cmd+K and insert a bulk of code in the middle. Also, I like the @Docs feature to bring latest documentation for popular libraries.

Weaknesses: Context falls apart in agentic mode; often loses track of previous instructions; requires excessive prompting Best for: Single-file optimizations and modifications, but not complex cross-file tasks

Claude Code - Declining quality since public beta

Strengths: Used to have superior reasoning and contextual understanding 3 months ago Weaknesses: Super expensive (like always), but recent updates have significantly degraded agentic capabilities; now requires much more hand-holding than before as it goes compleltely off base. Best for: Simple tasks that don't require deep contextual understanding Note: Most disappointing decline in quality - was previously much more capable. I spent $500 in Feb-Mar and thought it was worth.

Cline, Roo, and Aider - Conceptually interesting but practically limited

Strengths: Cline has good terminal integration; Roo offers interesting visualization; Aider has straightforward CLI Weaknesses: All three struggle with maintaining context; limited understanding of project structure; frequent need to repeat instructions Best for: Very simple, isolated coding tasks or experiments

Real-world Performance Differences The gap between these tools becomes most apparent when trying to implement complex features. For example, when asked to "add user authentication with email verification to my Express app":

Augment Code: Identified relevant files, added middleware, routes, and email service integration, then explained how the pieces fit together Windsurf/Cursor: Added authentication to single files I pointed at but needed explicit instructions for each additional component Others: Generally required file-by-file guidance with frequent context reminders

Conclusion If budget isn't a concern, Augment Code currently offers the most truly agentic experience, but still has a long way to go. For more budget-conscious developers, Windsurf slightly edges out Cursor for agentic capabilities, though both still require significant guidance for complex tasks.

Comments

SoMomentary•1y ago
I'm always surprised by people sleeping on GitHub Copilot. Is this because people truly don't find any value in it?
bv_dev•1y ago
I have used Github copilot since their beta release in 2023 and I don't find it anywhere near good these days. Automplete was good, but the industry has moved way beyond 2025. Copilot is slightly worse than Cursor which is itself a pretty average tool now. If you use truly agentic code generation, you won't be able to go back to Github Copilot.
SoMomentary•1y ago
You don't consider copilots agent mode to be agentic? I've had some pretty great results with agent mode + mcp to have it check it's own work.

Automated researchers can reliably mitigate alignment failures

https://www.anthropic.com/research/automated-researchers-mitigate-alignment-failures
1•EvgeniyZh•2m ago•0 comments

Show HN: Hikaru – image compression that hits an exact file size, no slider

https://hikarulabs.xyz
1•CFBL•3m ago•0 comments

The Ultimate Dream

https://blog.br11k.dev/2026-08-29-the-ultimate-dream
1•konovalov-nk•7m ago•0 comments

Git.kernel.org Is "Interesting" to Crawlers

https://people.kernel.org/monsieuricon/creepy-crawlies
1•sohkamyung•8m ago•1 comments

Show HN: Snippety – a native text expander for Mac and iPhone

https://snippety.app
1•snippety•8m ago•0 comments

We Ran Code Inside Fortune 500s Using Files They Published for AI Agents

https://medium.com/@alonhertz1/data-became-code-we-ran-code-inside-fortune-500s-using-files-they-...
2•nizbit•9m ago•1 comments

Does Computer Science Need Computers?

https://www.quantamagazine.org/does-computer-science-need-computers-20260828/
1•rbanffy•10m ago•0 comments

You Know GDPR Is Good Based on Who Hates It

https://matduggan.com/you-know-gdpr-is-good-based-on-who-hates-it/
1•latexr•13m ago•0 comments

Making $2.7M a Year Selling Ice in NYC [video]

https://www.youtube.com/watch?v=iGejCONYVjA
1•binyu•13m ago•0 comments

I ran a little experiment: could I make agent memory deterministic

https://www.lorekit.io/blog/agent-memory-data-model
1•mthines•15m ago•0 comments

Adversarial Interoperability

https://www.eff.org/deeplinks/2019/10/adversarial-interoperability
1•orphereus•20m ago•0 comments

"Lead through code" – engineering managers getting squeezed

https://www.youtube.com/watch?v=MPVmV-6O5to
1•lifeisstillgood•20m ago•0 comments

Show HN: Player vs. Computer

https://github.com/Rubinoslaw/Player-vs-Computer/
1•PolishDevelop22•23m ago•0 comments

LuaDB: Lightweight, embeddable, zero-dependency RDBMS written 100% in pure Lua

https://github.com/jncastilho/luadb
1•thunderbong•24m ago•0 comments

How bicycle coaster brakes work (2018)

https://www.dougbarnesauthor.com/2018/06/how-bicycle-coaster-brakes-work.html
1•Vedor•26m ago•0 comments

Rhine Group: Draghi, Collison unite experts for EU competitiveness

https://en.ilsole24ore.com/art/draghi-and-collison-stripe-set-to-boost-growth-in-europe-the-rhine...
2•Lucasoato•32m ago•0 comments

InferenceFS: Never worry about data again (Again)

https://github.com/philipl/inferencefs/
1•linkdd•33m ago•0 comments

Breakscale: Distributed Systems Design Simulator

https://github.com/xevrion/breakscale
1•handfuloflight•35m ago•0 comments

Muraqib, free nightly QA that lets Claude open a fix PR when tests fail

https://github.com/holistis/muraqib
1•Bugintelligence•48m ago•0 comments

Show HN: AgentBridge – Let one AI think while another AI writes the code

https://github.com/IndexFlowing/AgentBridge
2•mandarinclips•57m ago•1 comments

Nepal's flood damage may cost up to a tenth of economy to rebuild

https://www.reuters.com/world/china/nepal-needs-least-4-billion-rebuilding-finance-minister-says-...
2•geox•59m ago•0 comments

Pentagon's blacklisting of Anthropic was unlawful, US judge rules

https://www.theguardian.com/technology/2026/aug/28/us-court-rules-pentagon-anthropic-ban-illegal-...
2•vinni2•59m ago•0 comments

Walkable Cities

https://nartier.itch.io/walkable-cities
2•tehsauce•1h ago•0 comments

Some GitHub bounty repos are honeypots that farm free work from AI agents

https://oactodev.github.io/ninety-quid/report/
1•ninetyquid•1h ago•0 comments

Five ways 3D printing is improving lab work

https://www.nature.com/articles/d41586-026-02535-z
2•zeristor•1h ago•0 comments

Books Turned into Podcast Series

https://bookdialogues.com/home
1•jasmeet0817•1h ago•1 comments

Researcher Tricked Claude, Codex and Hermes into Running Malware

https://startupfortune.com/researcher-alon-hertz-tricked-claude-codex-and-hermes-into-running-mal...
12•CuriousLLM•1h ago•0 comments

SynapticChain: Single-wallet concurrency record (5,307 TPS from 1 key in Rust)

https://explorer.synapticchain.xyz
1•abdulshabazz•1h ago•0 comments

I built a financial visibility layer for multi-model AI applications

https://credlier.com/
1•dcriglerd•1h ago•0 comments

Show HN: Kith – AI clinical notes for therapists from ambient session audio

https://www.kith.space
1•kithmedai•1h ago•0 comments