# Context and Verification Are the New Coding-Agent Bottlenecks *By Coding Agents Alpha Tracker • July 18, 2026* Today’s most practical pattern is to design coding agents as resumable, verifiable loops: keep state outside the prompt, reset context before it degrades, and make commits pass deterministic gates. Also covered: Kimi K3 field tests, GPT-5.6 production comparisons, Grok Build open-sourcing, and agent-secret infrastructure. ## 🔥 TOP SIGNAL **Long-running agents are becoming useful enough that context and verification—not raw model capability—are the practical bottlenecks.** Theo’s Kimi K3 migration ran for more than three hours and completed 122 tasks from a short prompt before hitting its context limit; RALPH creator Geoffrey Huntley’s countermeasure is deliberately simple: preserve state in the filesystem, recycle the context window, and restart rather than letting an agent flail. [^1][^2] ## ⚡ TRY THIS - **Run a resettable loop, not one giant chat.** Put the desired end state in a prompt, keep progress and artifacts in files, then run a simple `while true` loop that re-reads those files each iteration. Huntley recommends treating roughly 100k tokens as the comfortable zone for hard work; if the model starts applying hacks or misidentifying stale tests, stop and begin a fresh context. [^2] - **Make the agent earn its commit.** Encode architecture and domain constraints in pre-commit hooks, including rules such as which boundaries may not depend on each other. Pair that feedback with static checks, linters, and semantic verification; Huntley’s point is that agents do not mind the friction, so the loop can be prevented from closing until requirements pass. [^2] - **Use vision as an explicit UI feedback loop.** Give the agent a reference screenshot, then require a repeatable cycle: launch the app, capture a screenshot, inspect it, edit, and re-check. Theo watched Kimi K3 follow this pattern; when browser behavior consumed a one-time authentication token, one clear steering instruction let it continue refining and report/clean up its mistake. [^1] - **Quarantine computer-use agents in VMs.** If an agent can drive a browser and desktop UI, run it in a VM so it cannot steal focus from your working machine. Peter Steinberger uses this with Codex while it handles GitHub interactions. [^3] ## 📡 WHAT SHIPPED - **Kimi K3 in agent harnesses:** Theo used the open-weight model through OpenCode and T3 Code for long migrations, phase-specific subagent workflows, security-audit discovery plus 25 verification agents, and screenshot-backed UI work. Caveat: the subscription setup he tested did **not** expose the full advertised context capacity and hit rate limits quickly under heavy use; Armin Ronacher separately warns that current maximum-thinking restrictions make K3 a poor fit for some basic tasks. [^1][^4] - **GPT-5.6 field reports are mixed by task, not benchmark.** DHH says GPT-5.6 Sol fixed an icon-alignment issue in 4m45s after GPT-5.5, Opus 4.8, and Gemini failed on that case. For code review, Steinberger reports 5.6 Terra high made his `clawsweeper` bot about 40% faster with negligible quality loss versus its prior setup, while Terra high outperformed Sol low for his own review use case—an explicit reminder to evaluate on your workload. [^5][^6][^7] - **Grok Build is now Apache 2.0 open source.** After backlash over directory uploads, xAI released its Rust-based coding-agent CLI; Simon Willison found 844,530 lines of Rust, visible system/subagent prompts, ports of Codex and OpenCode-style tools, and disabled remnants of GCS-upload code. Study it as a large agent harness, but the privacy incident is part of the operational context. [Repository](https://github.com/xai-org/grok-build) [^8] - **Treg:** Jason Zhou announced an open-source, self-hostable skill and secret registry that bundles skills with a CLI/endpoint while injecting authentication server-side, so agents do not hold keys; it also logs calls per agent and user. [Repository](http://github.com/superdesigndev/tools-registry) [^9][^10] - **Deep Agents:** LangChain’s harness packages planning, subagents, filesystem-backed context, sandboxes, and automatic compaction around 80% of the context window for long-running work. Its useful abstraction is progressive disclosure: offload large tool outputs and load files or skills only when needed. [^11] ## 🎬 GO DEEPER - **12:25–13:35 — Kimi K3’s 122-task migration run.** Watch the practical long-horizon test: a short prompt, a real codebase migration, then an abrupt context ceiling. It is a useful demonstration of both capability and the need for resumable state. [![Kimi K3 is the best model ever made (sometimes)](https://img.youtube.com/vi/Q4LoxsIwriA/hqdefault.jpg)](https://youtube.com/watch?v=Q4LoxsIwriA&t=745) *Kimi K3 is the best model ever made (sometimes) (12:25)* - **6:30–10:07 — Use the filesystem as your context layer.** The Deep Agents walkthrough explains why large tool outputs should live outside the active prompt, and how an agent can fetch them on demand rather than repeatedly summarizing away detail. [![Deep Agents Explained](https://img.youtube.com/vi/GbzEDgcuGJU/hqdefault.jpg)](https://youtube.com/watch?v=GbzEDgcuGJU&t=389) *Deep Agents Explained (6:29)* - **45:11–45:31 — RALPH reduced to its primitive.** Huntley’s short explanation of the `while true` + `cat` pattern is worth watching before building a more elaborate orchestration framework. [![The Great Loops Debate — Dex Horthy, Geoff Huntley, Ian Livingstone, Greg Pstrucha, @insecure-agents](https://img.youtube.com/vi/c35YoMdnI78/hqdefault.jpg)](https://youtube.com/watch?v=c35YoMdnI78&t=2713) *The Great Loops Debate — Dex Horthy, Geoff Huntley, Ian Livingstone, Greg Pstrucha, @insecure-agents (45:13)* - **Repository to study: Grok Build.** Its exposed prompt templates and implementations of familiar coding-agent tools make it a concrete codebase for examining how a large terminal-agent harness is assembled. [Open the repo](https://github.com/xai-org/grok-build) [^8] *Editorial take: the high-leverage upgrade is a resumable loop with hard verification gates—model improvements matter, but unattended context drift still decides whether long runs produce usable code.* [^2][^1] --- ### Sources [^1]: [Kimi K3 is the best model ever made \(sometimes\)](https://www.youtube.com/watch?v=Q4LoxsIwriA) [^2]: [The Great Loops Debate — Dex Horthy, Geoff Huntley, Ian Livingstone, Greg Pstrucha, @insecure-agents](https://www.youtube.com/watch?v=c35YoMdnI78) [^3]: [𝕏 post by @steipete](https://x.com/steipete/status/2078318731785359634) [^4]: [𝕏 post by @mitsuhiko](https://x.com/mitsuhiko/status/2078109495620771941) [^5]: [𝕏 post by @dhh](https://x.com/dhh/status/2078252859834167656) [^6]: [𝕏 post by @steipete](https://x.com/steipete/status/2078236791329657017) [^7]: [𝕏 post by @steipete](https://x.com/steipete/status/2078252386376929706) [^8]: [Kimi K3, and what we can still learn from the pelican benchmark](https://simonw.substack.com/p/kimi-k3-and-what-we-can-still-learn) [^9]: [𝕏 post by @jasonzhou1993](https://x.com/jasonzhou1993/status/2078081781786423462) [^10]: [𝕏 post by @jasonzhou1993](https://x.com/jasonzhou1993/status/2078084269747183753) [^11]: [Deep Agents Explained](https://www.youtube.com/watch?v=GbzEDgcuGJU)