What engineers are actually debating

The recurring theme on HN in 2026: agent workloads made the cost curve non-linear, and teams are realising per-token pricing scales in ways finance never signed up for. The debate has moved from "which model is cheapest" to "why is our agent quadratic".

• Expensively Quadratic: The LLM Agent Cost CurveHN · 131 pts · 81 comments · 2026

  Our take: This is the shift we see in every cost audit: an agent re-reads its whole context every step, so cost grows with the square of the task, not linearly. Routing and context management — not a cheaper model — are what flatten it.

• Why current LLM costs are not sustainableHN · 116 pts · 196 comments · 2026

  Our take: The 196-comment thread captures the anxiety accurately. Our read: sustainability at the app layer is an engineering problem you can solve now — caching, batching, routing — regardless of where provider pricing lands.

• We decreased our LLM costs with OpusHN · 106 pts · 31 comments · 2026

  Our take: Counterintuitive but real: a more capable model can be cheaper per finished task if it needs fewer retries and shorter loops. Cost per successful task, by route, is the number that matters — not the headline per-token price.

MCP went from a launch post to an industry standard fast. The community conversation has matured from "what is it" to governance, security, and the sprawl of integrations — which is exactly the gap between a demo server and one a team can trust.

• Model Context ProtocolHN · 872 pts · 258 comments · 2024

  Our take: The original launch thread. Worth reading for how quickly the room grasped that the value is a reusable, governed integration layer — not another chatbot plugin.

• Donating MCP and establishing the Agentic AI FoundationHN · 288 pts · 145 comments · 2025

  Our take: Standardisation and neutral governance are why MCP is a safe thing to build on now. That maturity is the signal to invest in internal MCP servers rather than wait.

• Show HN: mcp-agent – build effective agents with MCPHN · 80 pts · 28 comments · 2025

  Our take: The ecosystem is full of quick community servers like this. They are great for prototyping and exactly why production still needs auth, scoping, and audit logs layered on top.

The most heated developer topic of 2026, by comment volume. The pendulum thread — "after two years of vibecoding, I’m back to writing by hand" — and its counterpoints map the real tension: agents are productive and dangerous in the same breath.

• After two years of vibecoding, I’m back to writing by handHN · 865 pts · 634 comments · 2026

  Our take: The backlash is not anti-AI — it is anti-unreviewed-AI. Read it as a demand for the guardrails (tests, review hooks, golden paths) that make agent output trustworthy, which is what we build.

• Vibe coding and agentic engineering are getting closer than I’d likeHN · 787 pts · 885 comments · 2026

  Our take: The 885-comment thread is really about the line between a prototype and a production system. That line is exactly the productionization work — the "80% problem" made concrete.

• OpenCode – open source AI coding agentHN · 1274 pts · 618 comments · 2026

  Our take: Tooling is commoditising fast. The differentiator is no longer which agent you run but whether your codebase and workflow are ready for it — which is a readiness problem, not a tool choice.

Evals are where the community grew up. AccountingBench — evaluating LLMs on real long-horizon business tasks — landed hard because it measured what actually matters: can the model do the job over time, not ace a benchmark once.

• AccountingBench: evaluating LLMs on real long-horizon business tasksHN · 534 pts · 149 comments · 2025

  Our take: The lesson we take into every eval build: measure the real task over a real horizon, not a synthetic one-shot. Long-horizon degradation is invisible to most benchmarks and lethal in production.

• RouteLLM: a framework for serving and evaluating LLM routersHN · 244 pts · 36 comments · 2024

  Our take: Routing and evals are the same discipline: you can only route down a tier because an eval proves quality held. This thread is a good primer on treating routing as measurable.

• Task-specific LLM evals that do and don’t workHN · 182 pts · 46 comments · 2024

  Our take: Matches our experience exactly — generic eval suites underperform; the ones that work are narrow, task-specific, and owned by the team shipping the route.

Every few months a high-profile indirect prompt-injection exfiltration lands on HN’s front page — Slack AI, Google Bard, Antigravity. The pattern never changes: an agent that reads untrusted content and holds tools is the vulnerability, and no prompt fixes it.

• Google Antigravity exfiltrates data via indirect prompt injectionHN · 768 pts · 215 comments · 2025

  Our take: The newest entry in a long line. It reinforces our architectural stance: defence is least-privilege tools, audited access, and human gates on high-impact actions — not a cleverer system prompt.

• Data exfiltration from Slack AI via indirect prompt injectionHN · 604 pts · 174 comments · 2024

  Our take: A canonical case study. If an agent can read untrusted text and reach sensitive data in the same context, injection is a when, not an if. Separate those capabilities.

• Prompt injection: what’s the worst that can happen?HN · 462 pts · 203 comments · 2023

  Our take: Still the best framing of the threat model. We point clients here when they think injection is a theoretical concern rather than an architectural one.

Claude Code and its peers generate enormous, often critical, discussion — leaks, refusals, capability debates. The heat is a proxy for how much teams now depend on these tools, and how little they’ve standardised their use.

• Claude 3.7 Sonnet and Claude CodeHN · 2127 pts · 963 comments · 2025

  Our take: The launch that made terminal-native agents mainstream. The comment volume is the tell: this became core infrastructure for a lot of teams overnight, mostly without a rollout plan.

• Claude Code is unusable for complex engineering tasks (with recent model)HN · 1364 pts · 753 comments · 2026

  Our take: Threads like this are less about the tool and more about environment fit. The same agent thrives or thrashes depending on tests, boundaries, and context — which is the readiness work.

• Crush: glamourous AI coding agent for your favourite terminalHN · 367 pts · 235 comments · 2025

  Our take: New entrants arrive constantly. The strategic takeaway for a team: standardise the workflow (commands, rules, MCP access) so you can swap the underlying agent without re-training everyone.