# Make Coding Agents Verifiable Before You Make Them Autonomous *By Coding Agents Alpha Tracker • July 17, 2026* Today’s practical lesson is how to design a control plane around coding agents: centralize signals, validate outcomes, codify orchestration, and preserve human judgment at the highest-leverage points. Also: managed-agent updates from Gemini, a Deep Agents Code setup, and a multi-model Devin workflow. ## 🔥 TOP SIGNAL **The agent control plane is where the real work is moving.** Factory’s Eno Reyes says humans intervene most heavily on whether a proposed change and its plan/spec are actually the right ones; agents can then handle QA, code review, and security analysis. [^1] Armin Ronacher’s counterweight is blunt: remove every bit of friction and teams risk output they cannot understand or maintain—so keep review and operational checks while automating the toil around them. [^2] ## ⚡ TRY THIS - **Build a signal-to-plan inbox before you build an auto-merge loop.** Instrument customer Slack, GitHub issues, internal product discussions, telemetry, and other feedback into one pipeline. Have agents triage against a concise product-prioritization document, then make a human explicitly approve or reshape the resulting plan before agents implement, test, review, and analyze security. Factory keeps this guidance under roughly 2,000 lines. [^1] - **Turn a large task into a “mission” with validators.** Before implementation, force the agent to enumerate concrete outcomes and attach a validator to each: deterministic checks where possible, an LLM-based verifier only where necessary. Keep the loop running until every validation passes; run an agent-readiness check first. This is Factory’s approach for decomposing large engineering jobs. [^1] - **Make orchestration code, not emergent behavior.** Theo’s Claude Code workflow has the model write a JavaScript workflow file up front—stages, prompts, and sub-agents—then execute it top-to-bottom. Put the workflow convention in your global `agents.md` or system prompt so it becomes reusable; Theo reports these bounded workflows end cleanly and used about one-quarter of the tokens for comparable tasks in his testing. [^3] - **Treat permissions as a workflow decision.** Keep autonomous agents to local, reversible actions such as edits and tests; require confirmation for destructive, hard-to-reverse, shared, or externally visible actions. This is not theoretical: reported Codex file deletions most commonly involved full-access mode without sandboxing or auto-review, followed by a mistaken `$HOME` deletion during temporary-directory setup. [^3][^4] ## 📡 WHAT SHIPPED - **Gemini API managed agents:** Google added cost controls, a free tier, and initial triggers for scheduled agent tasks. [Related announcement](https://x.com/GoogleAIStudio/status/2077801843720093867?s=20). [^5] - **Deep Agents Code + NVIDIA Nemotron 3 Ultra:** LangChain highlighted a terminal-agent setup via Baseten with skills, sub-agents, MCP support, and LangSmith tracing; the cited model configuration is 550B parameters and up to 300 tokens/sec. [^6] - **Devin’s current multi-agent workflow:** Riley Brown’s hands-on walkthrough shows its Sessions/Kanban view for concurrent agents and per-agent model selection across SWE 1.7, Claude Fable 5, GPT-5.6 Luna, and GLM 5.2 High. The practical takeaway is less the model menu than the workflow: queue work, preview in-app, iterate with screenshots, and deploy through connected GitHub/Vercel tooling. [^7] - **Model routing is becoming a product control, not a developer habit.** Factory’s routing exposes an admin-set quality/cost slider and learns from a company’s task distribution; the company says typical savings are 30–50%, with the explicit tradeoff being quality. [^1] ## 🎬 GO DEEPER - **4:40–5:44 — Theo on bounded, code-defined subagent workflows.** A concise explanation of why a workflow file can be easier to control—and terminate—than agents spawning subagents opportunistically. [^3] [![I need you to hear me out (it’s REALLY good)](https://img.youtube.com/vi/Noo0NWD0gHU/hqdefault.jpg)](https://youtube.com/watch?v=Noo0NWD0gHU&t=279) *I need you to hear me out (it’s REALLY good) (4:39)* - **58:50–60:47 — Factory’s “missions” model.** Watch the outcome/validator framing: define what success means, select deterministic or LLM validation, and do not let the agent declare completion prematurely. [^1] [![The best AI agents need more humans than you think](https://img.youtube.com/vi/HbUznYhKFOc/hqdefault.jpg)](https://youtube.com/watch?v=HbUznYhKFOc&t=3473) *The best AI agents need more humans than you think (57:53)* - **26:39–29:25 — Armin Ronacher on guardrails that preserve human judgment.** Concrete examples include linting architectural constraints, automatically fixing mechanical issues, and escalating migrations or dependency changes for human review. [^2] [![A Year of Agents | Armin Ronacher | CodeCrafts 2026](https://img.youtube.com/vi/u_k9cwDNPcM/hqdefault.jpg)](https://youtube.com/watch?v=u_k9cwDNPcM&t=1599) *A Year of Agents | Armin Ronacher | CodeCrafts 2026 (26:39)* *Editorial take: the winning agent setup is not maximum autonomy—it is a system that makes plans reviewable, outcomes verifiable, and risky actions hard to perform by accident.* --- ### Sources [^1]: [The best AI agents need more humans than you think](https://www.youtube.com/watch?v=HbUznYhKFOc) [^2]: [A Year of Agents | Armin Ronacher | CodeCrafts 2026](https://www.youtube.com/watch?v=u_k9cwDNPcM) [^3]: [I need you to hear me out \(it’s REALLY good\)](https://www.youtube.com/watch?v=Noo0NWD0gHU) [^4]: [Quoting Thibault Sottiaux](https://simonwillison.net/2026/Jul/16/bad-codex-bug) [^5]: [𝕏 post by @OfficialLoganK](https://x.com/OfficialLoganK/status/2077810190179762366) [^6]: [𝕏 post by @LangChain](https://x.com/LangChain/status/2077770240985887082) [^7]: [Devin AI: The Full Beginner’s Guide \(Better Than Claude Code?\)](https://www.youtube.com/watch?v=Vyyrvna-hUY)