A significant evolution in AI-assisted development is underway. The industry is moving beyond autocomplete-style code suggestions toward AI agents capable of sustaining work across hours, days, or even weeks — handling complex, multi-step engineering projects with minimal human intervention between checkpoints.
The concept, explored in a long-form analysis by Addy Osmani republished on O'Reilly Radar, describes "long-running agents" as a new class of AI system that maintains persistent state, recovers from failures, generates structured artifacts, and resumes work across multiple sessions and sandboxes. Unlike traditional AI coding assistants bounded by a single context window, these agents function as autonomous workers pursuing defined goals over extended timeframes.
Beyond Autocomplete: A New Tier of Automation
The practical implications for engineering teams are substantial. While current AI coding tools excel at line-level suggestions and short-function generation, long-running agents are designed to tackle bounded but ambitious workflows: migrating legacy codebases, conducting extended security audits, orchestrating large-scale refactoring, or scaffolding entire subsystems from architectural specifications.
As Osmani describes, these agents produce structured, machine-readable outputs — code, logs, decision records — that serve as auditable project memory. This shifts the role of AI from a reactive assistant responding to prompts into a goal-directed worker that generates organized, verifiable deliverables.
For technology teams operating in complex environments — particularly those managing bilingual or multilingual legacy systems, older monolithic architectures, and diverse regulatory requirements — this capability could prove especially valuable. Modernizing such systems is typically labour-intensive, poorly documented, and resistant to incremental automation. An agent capable of sustained, methodical work across a multi-week migration could meaningfully reduce timelines and improve consistency, provided adequate human oversight is maintained.
Architectural Foundations: What Teams Need Before Adoption
The readiness of long-running agents depends on three technical pillars, according to the analysis.
Persistent state management allows an agent to pause and resume without losing context — a critical requirement when tasks span days. Robust error recovery and sandboxing ensure that failures in one session do not corrupt progress or compromise security. Traceability infrastructure — dashboards, logging systems, and decision-path visualizations — enables human reviewers to audit thousands of autonomous decisions without manually inspecting every line of output.
For organizations considering adoption, the investment required extends beyond the agent itself. Teams need durable artifact storage designed for both human and machine review, checkpointing and rollback mechanisms, and clear access-control policies for agents with persistent permissions.
Rethinking How Humans and AI Collaborate
Perhaps the most consequential shift is organisational rather than technical. Long-running agents require a different model of human oversight — one centred on periodic review and strategic steering rather than real-time monitoring. This means rethinking team workflows: establishing intervention protocols, defining escalation triggers, and designing feedback loops that let engineers course-correct without becoming bottlenecks.
Osmani's analysis recommends starting with well-defined, high-value projects where agent output can be easily verified — not open-ended creative work. The goal is iterative trust-building: demonstrate reliability on bounded tasks before expanding an agent's autonomy.
Open Questions Remain
Significant uncertainties persist. There is no consensus on the most effective interfaces for humans to steer week-long agent tasks. Quantifying the return on investment — and determining trust thresholds for expanding autonomy — lacks established frameworks. And whether cross-platform standards for agent state portability and artifact schemas will emerge, or whether the ecosystem will fragment along vendor lines, remains an open question.
The Strategic View
For technology leaders and engineering teams, the emergence of long-running agents signals a maturation of the AI development tooling landscape. The conversation is shifting from "can AI write this function?" to "can AI manage this migration?" That progression demands new infrastructure, new governance models, and a willingness to rethink how complex technical work is organised and overseen.
The potential gains are clear, but the technology is early. Teams looking to position themselves should act now: pilot agents on bounded, well-documented tasks such as incremental legacy migration, invest in observability infrastructure that makes autonomous decisions auditable, and build the organisational muscle for periodic-review workflows before scaling agent autonomy further.
AI 輔助開發正在經歷一場重大演變。業界正超越自動補全式的程式碼建議，轉向能夠持續工作數小時、數天甚至數週的 AI 代理——這些代理能在檢查點之間以最少的人工干預，處理複雜、多步驟的工程專案。
這一概念在 Addy Osmani 撰寫並於 O'Reilly Radar 轉載的長篇分析中有探討，其將「長時間運作的代理」描述為一類新型的 AI 系統，它能維護持久狀態、從故障中恢復、生成結構化成果，並跨多個工作階段和沙盒恢復工作。與受單一上下文窗口限制的傳統 AI 編程助手不同，這些代理的功能如同自主工作的工人，在更長的時間跨度內追求既定目標。
超越自動補全：自動化的新層級
對工程團隊而言，其實際意義重大。雖然現有的 AI 編程工具擅長行級建議和短函數生成，但長時間運作的代理旨在處理有界限但雄心勃勃的工作流程：遷移遺留程式碼庫、進行擴展的安全審計、編排大規模重構，或根據架構規格搭建整個子系統的骨架。
正如 Osmani 所述，這些代理會產生結構化、機器可讀的輸出——程式碼、日誌、決策記錄——這些輸出可作為可審計的專案記憶。這使 AI 的角色從對提示做出回應的被動助手，轉變為生成有組織、可驗證交付成果的目標導向工作者。
對於在複雜環境中運作的技術團隊——特別是那些管理雙語或多語種遺留系統、較舊的單一整體架構以及多樣化監管要求的團隊——這項能力可能尤為寶貴。對這類系統進行現代化改造通常是勞動密集型、文件記錄不足，且對漸進式自動化有抵抗性的。一個能夠在數週遷移工作中持續、有條不紊工作的代理，如果能在足夠的人類監督下進行，有望顯著縮短時間線並提高一致性。
架構基礎：團隊採用前所需準備
根據分析，長時間運作的代理的就緒程度取決於三大技術支柱。
持久狀態管理允許代理在不丟失上下文的情況下暫停和恢復——這是任務跨越數天時的關鍵要求。穩健的錯誤恢復與沙盒化確保單一工作階段的失敗不會破壞進度或危及安全。可追溯性基礎設施——儀錶板、日誌系統和決策路徑可視化——使人類審查者能夠審計數千個自主決策，而無需手動檢查每一行輸出。
對於考慮採用的組織而言，所需的投資不僅僅是代理本身。團隊需要為人類和機器審查而設計的持久成果儲存、檢查點和回滾機制，以及針對擁有持久權限的代理的清晰存取控制策略。
重新思考人類與 AI 的協作方式
或許最具深遠影響的轉變是組織層面的，而非技術層面的。長時間運作的代理需要一種不同的人類監督模式——一種以定期審查和策略指導為中心，而非即時監控的模式。這意味著需要重新思考團隊工作流程：建立干預協議、定義升級觸發器，以及設計讓工程師能在不成為瓶頸的情況下進行修正的反饋迴路。
Osmani 的分析建議從定義明確、價值高的專案開始，在這些專案中代理的輸出易於驗證——而非開放式的創意工作。目標是逐步建立信任：先在有界限的任務上展示可靠性，然後再擴展代理的自主權。
懸而未決的問題
重大不確定性仍然存在。對於人類引導長達一週的代理任務最有效的介面，目前尚無共識。量化投資回報率——以及確定擴展自主權的信任閾值——缺乏既定框架。至於代理狀態可移植性和成果架構的跨平台標準是否會出現，或者生態系統是否會沿供應商路線碎片化，這仍然是一個開放的問題。
策略視角
對於技術領導者和工程團隊而言，長時間運作的代理的出現標誌著 AI 開發工具生態的成熟。對話正從「AI 能寫這個函數嗎？」轉向「AI 能管理這次遷移嗎？」。這種進步需要新的基礎設施、新的治理模式，以及重新思考複雜技術工作如何被組織和監督的意願。
潛在的收益是明確的，但這項技術仍處於早期階段。希望搶佔先機的團隊應立即行動：在界限明確、文件完善的任務（如漸進式遺留系統遷移）上試點代理；投資於使自主決策可審計的可觀察性基礎設施；並在進一步擴展代理自主權之前，建立定期審查工作流程的組織能力。