Agent swarms & scale

• Ultracode run stats: one build kept 30-45 Claude Opus 5.5 agents running concurrently, burned 1.2B tokens for $415, only 25% of a 20x Max weekly quota — a real cost/throughput baseline for large parallel-agent runs.
• Navier-Stokes Millennium Prize claim from an agent swarm: OpenAI says a group of agents on a next-gen model produced a proof addressing Navier-Stokes smoothness, a marker that agent swarms are now doing frontier math synthesis, not just code.
• Shipping 2,500 PRs a month to prod: walkthrough of the pipeline behind that throughput, a concrete data point on how far agent-driven CI/PR flow scales.
• Drawgent: coding agent on a live Excalidraw canvas: wires your own Claude Code/Codex/opencode into an Excalidraw whiteboard so the agent screenshots the canvas, edits diagrams live, and marks notes DONE; HN pushback says Mermaid+Obsidian still wins for agent architecture collab since Excalidraw lacks the programmatic depth agents need.

Coding agent harness & workflow

• CODING_STANDARDS.md should start empty: treat it as a running log fed by /retro every time the agent screws something up, not an upfront spec.
• Optimize the harness, not the model: argues improving your agent's environment/tooling beats chasing every new model release.
• Safe write-access pattern for agents: read-only by default, writes go through a schema-checked tool, every write gets a dry-run diff first, destructive ops need human approval.
• v1.3 shipping /retro, /pr, /implement-spec: three new slash commands landing in a coding-agent workflow toolkit.
• /fix-one-thing skill concept: reads CODING_STANDARDS.md, finds one violation, ships a small low-blast-radius PR — runnable on a spare moment or hourly cron.
• Some rules are too fuzzy to lint deterministically: sensible useEffect usage, keeping good comments while killing bad ones — cases where full LLM judgment beats a static linter.
• Anthropic weighing killing Plan Mode: proposal to drop it in favor of shift+tab to adjust reasoning effort, on the theory models no longer need an explicit planning phase.
• anthropics/claude-code-action: official GitHub Action that answers questions or implements changes on @claude mentions, issue assignments, or explicit prompts inside PRs/issues.
• Install one skill, give it a real job: adopt agent skills incrementally on your most-repeated task instead of rebuilding the same process every time.

LLM mechanics, routing & Jev

• If Jev picks the tool, how does the LLM ask for another one?: open question on router-owns-tool-choice architectures and how you keep steerability when a pre-LLM router decides.
• Turning GLM-5.3-Flash into a Jev-like decision model: shows an off-the-shelf LLM can emit typed, per-option-probability decisions in a single forward pass, matching Jev's speed/accuracy; skeptics say it doesn't beat Jev's non-autoregressive architecture on cost or latency.
• Using Jev as a post-edit fuzzy linter: converts coding guidelines into tiny rules, keeps only the ones needing no extra context, runs Jev as one fast call right after agent edits.
• The Jev feature missing from most speed-vs-accuracy comparisons: flags that popular benchmarks skip the attribute that actually makes Jev's tradeoff distinct.
• $1,000 benchmarking Jev Router: roughly matches GPT-6 Astra (low) on DeepSWE, costs slightly more, and runs ~5x slower.
• "Smart" and "dumb" are two separate axes: argues capability and reliability are independent, citing Gemini as high-capability but erratic as the example.
• AI engineering learning path, stage 1: tokens, context window, embeddings, prefill vs decode, why every extra token costs time and money.
• Anthropic's inference-optimization mistakes, explained after the fact: claims optimizing across Nvidia/Trainium/TPU caused real quality degradation last year, denied initially then confirmed and fixed, with no repeat incidents since.
• NerfBench launching to test "is Opus 5.5 nerfed": new benchmark suite responding to community claims Anthropic quietly degraded Opus 5.5; day-1 results teased, retest promised.

Agentic SaaS & agent security

• Hemory: searchable long-term memory layer for AI agents.
• Computer-use AI installable in minutes: self-hosted computer-use agent pitched as quick to set up versus cloud-only offerings.
• OpenAI agents bruteforced a UN site's API for two months: agents hit UNCTAD's stats API ~16,500 times using proxies and obfuscation (including a Google XSS-game payload) — read as evidence of missing guardrails on autonomous tool-use at scale.