12 talks, 6 themes, one fast way through the event — key takeaways, standout quotes, and every recording from Coding Agents Conference.
Worth noting
Argues general-purpose agents need a coding-agent-style harness (Deep Agents: virtual file system, planning, sub-agents, compaction) plus a tool runtime doing delegated per-user OAuth so agents act as users, not with over-privileged service tokens.
Quote
Doing work for the user is good. Doing work as the user is much better in many cases.
Worth noting
Dexter Horthy retires RPI for "CRISPI": splitting the 85-instruction mega-prompt into 7 smaller stages, since research shows LLMs reliably follow only ~150-200 instructions, and shifts review from long plans to lightweight design/outline docs plus reading actual code.
Quote
Frontier LLMs could only follow about 150 to 200 instructions with good consistency.
Worth noting
Kilo's 15 engineers each own one feature end-to-end, running 2–4 parallel agents, shipping 1–2 features weekly (up from one every 2–3 weeks) across 25 trillion tokens processed and 1.5M developers.
Quote
You've got to give trust to get trust... as you get deeper up that ladder, you need to open up the books and give more context.
Worth noting
Warp's founder argues local laptops can't scale multi-agent coding, and demos Oz, a cloud platform for launching, tracking, and team-sharing coding-agent runs across any harness.
Quote
I think it sometime in 2025 coding switched over from being you writing code by hand to writing by prompt.
Worth noting
A customer's deep-research agent (250k products, ~200 steps/100 LLM calls each) went from prototype to 2,000+ concurrent runs in an hour using replay logs, caching, and spot instances.
Quote
The problem isn't that agents fail, it's that recovering from failure is challenging without the full context of how infra, networking, logical, semantic layers all interact.
Worth noting
Fastino's upcoming Pioneer platform continuously runs agents that monitor inference logs, fine-tune, and re-evaluate deployed open-source models (Llama, Qwen, DeepSeek) to counter model drift and cut cost/latency.
Quote
What that looks in reality is that you end up with the situation in which you've deployed Llama and it's actually getting more accurate over time.
Worth noting
In a head-to-head eval on SWE-bench and TypeScript-Go bug fixes, agentic search (Claude Code's grep/find approach) beat a basic vector search 68% vs 60%, while costing far fewer tokens.
Quote
Evals are important because without them you're essentially making ship decisions based off of vibes, which is not good.
Worth noting
SWE-Bench Pro fought contamination with proprietary code and human-verified tasks; a new survey found issue resolution is only ~25% of engineering work, prompting a new, broader benchmark launching next.
Quote
We start seeing a trend of coding being the legs and hands for AI.
Worth noting
Pinterest cut LLM post-training from 4-6 weeks to ~1 week using Claude Code sub-agents; letting agents write their own code ("tool calling 2.0") cut token use by 50-70%.
Quote
We essentially giving the agent capability to write its own code to solve its question.
Worth noting
Cleric, an AI SRE deployed to dozens of customers, found learning agents need three things: easy correction, corrections that persist/compound/are visible, and continuous ambient context absorption.
Quote
Wrong answers are inevitable and accuracy is important but the agents that win will be the ones that make it easy to correct mistakes and not visibly repeat those mistakes later.
Worth noting
Offers three concrete steps for securing AI coding agents: minimally scope credentials, log every action via hooks, and automatically scan agent-generated code with tools like Semgrep.
Quote
Downscope your credentials before you start. It takes two minutes and can save you from a really bad day.
Worth noting
Databricks built an internal "coding agent gateway" (Isaac) unifying billing, observability and MCP token management across Claude Code, Codex, Cursor etc. for 2,000+ engineers, shifting bottlenecks to code review and CI.
Quote
You should absolutely use the best tool for your use case, even if that's not the one-size-fits-all solution that you have already.
