Hacker News discussion: Agentic Coding Is a Trap

Captures the practitioner response to Faye's thesis — including a sharp framing from turtleyacht that orchestrating an agent and writing code may not even be the same cognitive activity.

Same model, different harness: 52.8% to 66.5% on Terminal-Bench 2.0

@Mason Daugherty (@masondrxy) on X

A specific numeric data point on harness leverage. Fourteen points of benchmark performance from changing scaffolding around a frozen model is the kind of result that should change how teams allocate engineering time.

“the same model in a different harness can yield much different performance”

Agentic Coding is a Trap

larsfaye.com

A senior developer arguing — with citations from Anthropic's own research — that the orchestration workflow is eating the orchestrator. Worth engaging with directly because it names a specific mechanism: the skills that make you a good supervisor are the ones the supervision is d

“You cannot replace a deterministic system with a probabilistic one and expect zero ambiguity.”

A founder paid $8k for an AI-built healthcare MVP. Then the pilot clinic asked for a HIPAA BAA.

reddit.com

A field report on where agentic coding hits a wall that no model release will solve: domain knowledge that has to be in the architecture from day one. The mechanic of the failure — Cursor doesn't know what a BAA is — is sharper than any general critique of AI coding.

“Cursor doesn't know what a BAA is. The prompts never asked for it.”

Most of my Claude usage was on work that didn't need Claude. Cut my bill 60x on bulk tasks with a tiny side model.

reddit.com

A practical pattern any builder can apply tomorrow: a deny-list in CLAUDE.md and a cheap side-model for mechanical work. The 60x cost gap is real, and the negative-framing detail is the kind of thing you only learn from running this for three weeks.

“positive instructions got treated like suggestions, deny lists got treated like rules.”

Memtrace: rewind-and-replay context for Claude Code

github.com

A concrete attempt to give a coding agent something a senior engineer takes for granted: a time-aware view of the codebase. The architectural bet — let the compiler do the structural work, don't pay an LLM to re-derive it — is the kind of design decision the harness layer is conv

“You don't pay an LLM to re-derive what your compiler already knows.”

llama.cpp MTP support now in beta

github.com

Local inference keeps catching up to the production stack. For builders who run open weights on their own hardware, this is the kind of incremental release that materially changes throughput on a Tuesday.

“Between this and the maturing tensor-parallel support, expect most performance gaps between llama.cpp and vLLM, at least when it comes to token generation speeds, to be erased.”

Pushing a 5-Year-Old 6GB VRAM laptop to Its Limits: Qwen3.6-35B-A3B

reddit.com

A 35B-class model at usable speed on a five-year-old gaming laptop. The hardware floor for serious local inference keeps dropping, and the recipes are specific enough to copy.

DeepClaude: Claude Code agent loop with DeepSeek V4 Pro

alattaran

A working demonstration of the harness-as-durable-artifact thesis. Same CLI, same workflow, different model behind a compatible endpoint — and the fact that it's four lines of shell makes the lock-in argument from the Faye essay much more concrete.

“export ANTHROPIC_BASE_URL=https://api.deepseek.com/anthropic — that's the whole switch.”