Agentic AI · since 2021
Agentic Coding
AI that writes, reviews and operates software — from autocomplete (GitHub Copilot, 2021) to full-lifecycle agents that plan, edit, run tests and ship (Codex, Claude Code, Cursor, Devin). A multi-vendor field measured on SWE-bench; the 2026 enterprise wave is its maturity, not its origin.
46
events traced
46
source records
30 Jun 2026
first signal
9 Jul 2026
last activity
Who drove it
Key movements
2021origin
origin (GitHub Copilot)
2023★
category benchmark (SWE-bench)
4+↑
2026 Gartner: multi-vendor Leaders
The story, event by event
Every point below is traced to a real source — nothing on this page is invented.
29 Jun 2021impact 84
GitHub Copilot — AI coding begins
value: Copilot; metric: origin
10 Oct 2023impact 74
SWE-bench — the category benchmark
value: SWE-bench; metric: benchmark
12 Mar 2024impact 72
Devin — the first "AI software engineer"
value: autonomous; metric: inflection
22 May 2025impact 76
Terminal agents — Claude Code & Codex
value: terminal agents; metric: generation
6 Oct 2025impact 60
Introduction of CodeMender for Code Security
examples: identifying vulnerabilities, creating complex patches; enhancements: automated validation processes, advanced program analysis tools, improved root cause identification
18 Nov 2025impact 60
Introduction of Gemini 3 Pro
model name: Gemini 3 Pro; capabilities: agentic coding, vibe coding, multimodal understanding; leaderboard score: 1487
28 Jan 2026impact 40
Introduction of Upskill Tool for Coding Tasks
token usage: optimized for recurring tasks; performance improvement: 35%
12 Feb 2026impact 60
Introduction of OpenEnv Framework
OpenEnv framework enhances AI agent evaluation.
OpenEnv is an open-source framework from Meta and Hugging Face designed to address this challenge by standardizing how agents interact with real environments.
Calendar systems are deceptively complex.
Agents achieved close to 90% success on tasks with explicit calendar identifiers, but success dropped to roughly 40% when the same tasks were phrased using natural language descriptions.
Multi-step reasoning is the primary bottleneck.
Turing built a production-grade calendar management environment referred to as the Calendar Gym.
These challenges are not unique to scheduling and calendars.
The introduction of the OpenEnv framework marks a crucial step in enhancing the evaluation of AI agents, particularly in complex real-world scenarios.
13 Feb 2026impact 60
Agent Skill for Optimized CUDA Kernels
average speedup: 1.88, 1.94; bandwidth efficiency: 34.7, 22.3
9 Mar 2026impact 60
LeRobot v0.5.0 Release
features: Full support for Unitree G1 humanoid robots, New policies for improved inference, EnvHub for simulation management, Streaming video encoding, Support for OpenArm robot, CAN bus support for actuators, Enhanced dataset pipeline performance, Integration with NVIDIA IsaacLab-Arena, Modernized codebase requiring Python 3.12; research paper: Accepted at ICLR 2026
10 Mar 2026impact 60
Asynchronous Training Architectures in RL Libraries
key findings: Ray dominates orchestration (8/16 surveyed distributed computing libraries), The inference pool runs continuously, feeding completed rollouts into a buffer, MoE support is an increasingly important differentiator as the field moves toward sparse models, Staleness management handles data that was generated under an old policy, Eight of the thirteen libraries support pushing only the LoRA adapter deltas to the inference server
16 Apr 2026impact 60
Major update to Codex
features: operates a computer, over 90 additional plugins, re-uses conversation threads, updates for desktop app users; developer count: 3000000
23 Apr 2026impact 60
Codex Automation Enhancements
automation features: {'recurring_tasks': True, 'scheduled_tasks': True, 'proactive_management': True}
23 Apr 2026impact 60
Codex Enhancements
features: Manages tasks across files and tools, Accessible to non-developers, Connects to tools for automation; description: Codex's capabilities in task automation and management.
23 Apr 2026impact 60
Enhanced Codex Functionality with Plugins and Skills
skills: {'description': 'Skills function as playbooks for Codex to execute tasks.', 'rules_adherence': 'A skill helps Codex follow those rules without making you explain them every time.'}; plugins: {'description': 'Plugins enable Codex to access external tools and information.', 'technical_skill': 'Creating a new plugin usually requires more technical expertise than creating a skill.'}
7 May 2026impact 65
Simplex achieves significant productivity gains in software development with Codex and ChatGPT Enterprise
Simplex has significantly enhanced its software development productivity by integrating Codex and ChatGPT Enterprise, leading to substantial time reductions across various stages.
Codex has reduced screen development time by 70%.
Screen design time has been cut by 40% with the use of Codex.
Internal integration testing time has decreased by 17% due to Codex.
Simplex is also validating automated workflows that run Python scripts from Codex CLI, exploring non-linear development processes with upfront rules.
This demonstrates a concrete, measured impact of agentic AI tools on enterprise software development workflows, highlighting a shift towards AI-driven operating models and significant productivity gains.
8 May 2026impact 55
Secure Deployment & Governance for Coding Agents
OpenAI details the secure deployment and governance framework for its Codex coding agent, balancing developer productivity with robust security measures.
Security teams require governance over agent operations, including control over access, human approval requirements, system interactions, and telemetry for behavior explanation.
Codex is deployed with the principle of being productive within a bounded environment, allowing frictionless low-risk actions and requiring review for higher-risk actions.
Codex activity logs are made available through the OpenAI Compliance Platform for enterprise and educational customers, enhancing transparency and auditability.
This progression highlights the increasing maturity of agentic coding solutions, moving beyond pure capability to focus on enterprise-grade security, governance, and compliance, which are critical for broader adoption.
12 May 2026impact 75
AutoScout24 Group achieves 10x faster development with agentic coding
AutoScout24 Group significantly accelerated its engineering and innovation capabilities by implementing a dual-layer AI strategy, integrating ChatGPT for broad literacy and Codex for specialized builder roles.
AutoScout24 achieved approximately 10x faster development cycles, reducing timelines from weeks to days, and specifically from 2-3 weeks to 2-3 days for select projects.
The company enabled around 2,000 employees with AI tools, with about 1,000 builder roles utilizing Codex for tasks such as automated pull request reviews, large-scale refactoring, technical documentation, and post-incident analysis.
This adoption also improved code quality and consistency through automated reviews and reduced manual workload in documentation and pull request processes.
This demonstrates a strong case study for the measurable impact of enterprise-wide AI adoption, particularly for agentic coding tools, in driving significant improvements in development efficiency and innovation capacity.
12 May 2026impact 65
AI coding agents impact ML competitions and review processes
AI coding agents significantly impacted the Parameter Golf machine learning challenge, demonstrating their role in lowering barriers to entry and reshaping competition dynamics.
The Parameter Golf challenge observed widespread use of AI coding agents, which 'helped lower the cost of experimentation, made it easier for more people to participate, and changed the pace of the competition'.
The prevalence of AI agents also 'created new challenges for submission review, attribution, and scoring' for the competition organizers.
To manage these challenges, organizers 'developed an internal Codex-based triage bot to monitor new submissions and flag them for human review'.
Beyond technical use, AI agents, such as '@notapplica's', also 'became part of the community around the challenge', providing 'Live Updates' and explaining leaderboard approaches.
The integration of AI coding agents into competitive technical environments highlights their growing influence on developer workflows and the need for new strategies in managing and evaluating AI-assisted work.
12 May 2026impact 55
Codex automates financial reporting and analysis for finance teams
Codex is being utilized by finance teams to automate the creation of review-ready financial documents and reports, streamlining financial analysis processes.
Codex automates the transformation of financial data into reviewable outputs without requiring coding, enabling teams to efficiently produce structured documents.
It generates narrative documents that include source-backed financial numbers and identifies key variances, changes since forecast, risks, and CFO prep questions.
The system is designed to create CFO-ready reports that explain performance and key variances, allowing finance teams to focus more on analysis and decision-making.
This demonstrates the expanding practical application of agentic coding beyond core software development into specialized business functions, enhancing efficiency and strategic focus in financial operations.
13 May 2026impact 65
Codex Windows Sandbox Implementation
OpenAI successfully engineered a custom, multi-component sandbox solution for Codex on Windows, overcoming the operating system's lack of native isolation primitives for open-ended coding agent workflows.
When the Codex engineering team began work in September 2025, Codex for Windows lacked a sandbox, forcing users to choose between insecure or limited options, as Codex runs with real user permissions by default, posing a significant security risk.
Windows did not provide out-of-the-box isolation utilities comparable to Seatbelt on macOS or seccomp/bubblewrap on Linux, and existing Windows features like Windows Sandbox were unavailable on common SKUs or presented greater risks (e.g., MIC integrity labeling).
The team iterated through prototypes, starting with an 'unelevated sandbox' using Windows concepts like write-restricted tokens to avoid admin privileges, but faced limitations with desired tools like Windows Firewall requiring elevation.
The current implementation, the 'elevated sandbox,' requires admin permissions during setup to achieve the necessary isolation and security, balancing robust protection with user experience for coding agents.
This engineering effort demonstrates the critical need for custom security solutions when deploying advanced agentic AI systems on diverse platforms, highlighting the ongoing work required to ensure safe and effective real-world application of agentic capabilities.
14 May 2026impact 65
Codex mobile app integration for remote task management
Codex is now integrated into the ChatGPT mobile app, enabling remote task management and collaboration across devices.
Codex is integrated into the ChatGPT mobile app for remote work management, allowing users to stay informed while Codex operates across various environments like laptops and devboxes.
The widespread adoption of Codex is highlighted by its more than 4 million weekly users, underscoring the importance of these incremental advancements.
Technically, Codex utilizes a secure relay layer to maintain connectivity with trusted machines across devices without direct public internet exposure.
This expansion of Codex's accessibility and functionality through mobile integration signifies a growing trend towards ubiquitous agentic AI capabilities, making these tools more pervasive in daily workflows and enhancing remote productivity.
14 May 2026impact 65
Sea Limited adopts agentic coding workflows with Codex
Sea Limited has integrated the AI coding tool Codex across its developer organization, achieving 87% weekly active users and fundamentally shifting its software development paradigm towards agentic workflows.
Sea is rolling out Codex across its developer organisation, with our internal data showing 87% of users are weekly active users.
AI agents are increasingly operating within our CI/CD pipelines—reasoning through product requirements, autonomously proposing test-driven implementations, surfacing edge cases in distributed systems, and accelerating debugging loops.
The 'developer' evolves into a 'system orchestrator' who spends the bulk of their time on tasks such as product judgment, system design, and orchestrating AI-driven workflows.
This enterprise-wide adoption demonstrates the maturation of agentic coding tools beyond simple autocomplete, enabling a deeper transformation of engineering practices and developer roles.
15 May 2026impact 55
Codex automates data science deliverable drafting
Codex automates the initial drafting of data science analysis deliverables, enabling teams to accelerate their workflow.
Codex helps data science teams convert scattered inputs into analysis assets quickly.
It drafts deliverables from various sources like dashboards, metric definitions, and business context, including charts, caveats, and review questions.
This allows analysts to focus their judgment on validating evidence, testing caveats, and refining recommendations.
Suggested plugins for Codex include Google Drive, Spreadsheets, Slack, Gmail, and Documents, indicating its tool-calling capabilities.
This demonstrates the expanding application of agentic AI to automate complex, multi-step knowledge work beyond traditional software development, enhancing productivity in specialized domains like data science.
18 May 2026impact 65
OpenAI & Dell Partner for Enterprise Codex Deployment
OpenAI and Dell Technologies are collaborating to integrate Codex into enterprise environments.
OpenAI and Dell are collaborating to deploy Codex in enterprise environments.
Codex will integrate with the Dell AI Data Platform for enterprise data management.
The collaboration provides customers with a practical approach to deploying AI solutions.
This partnership signifies a move towards more practical and secure deployment of agentic coding tools within existing enterprise infrastructure, accelerating their adoption.
20 May 2026impact 78
Enterprise wave + Gartner Leaders quadrant
value: multi-vendor; metric: market
20 May 2026impact 65
Ramp leverages Codex with GPT-5.5 for accelerated code review and agentic tooling
Ramp engineers are using Codex with GPT-5.5 to significantly accelerate code review and develop advanced agentic tools, improving feedback cycles and code quality.
Ramp engineers are using Codex with GPT-5.5 to accelerate code review and develop internal agentic tooling, helping teams get substantive pull request feedback in minutes instead of hours.
Codex code review catches things that I miss and that other engineers miss and that other AI code reviewers definitely miss.
Ray is also using Codex to support the development of On-Call Assistant, an agentic tool that takes on most of the burden for Ramp engineers during on-call rotations.
Engineers are going to become orchestrators. The skill is no longer writing every line of code yourself. It’s knowing how to direct AI tools like Codex, when to trust them, and when to push back.
This demonstrates the increasing practical utility and adoption of AI in software development workflows, shifting the role of engineers towards orchestration and leveraging AI for higher quality and efficiency.
22 May 2026impact 75
Virgin Atlantic deploys Codex for app development and legacy refactoring
Virgin Atlantic significantly improved software development velocity and quality by integrating Codex into its engineering and data workflows.
Codex led to a 78-80% codebase size reduction on legacy refactors and reduced legacy code refactoring time from two weeks to 30 minutes.
The revamped mobile app achieved near-complete unit test coverage (~100%) and zero P1 defects at launch, enabling a successful release before the high-risk Christmas travel rush.
Codex also empowered analyst teams to prototype internal applications directly against the company's data warehouse, accelerating data-related development.
This case study demonstrates the tangible benefits of agentic coding in enterprise settings, showcasing significant gains in efficiency, quality, and broader accessibility of development tools beyond traditional engineers.
22 May 2026impact 65
Codex recognized as Gartner Leader for Enterprise AI Coding Agents
OpenAI's Codex has been recognized as a leader in the Gartner Magic Quadrant for Enterprise AI Coding Agents, highlighting its advanced capabilities and significant enterprise adoption.
OpenAI's Codex was named a Leader in the Gartner Magic Quadrant for Enterprise AI Coding Agents, validating its progress in supporting large-scale enterprise deployments ("OpenAI has been recognized as a Leader in the Gartner® Magic Quadrant™ for Enterprise AI Coding Agents.").
The evaluation noted Codex's strengths in agentic software development, enterprise governance, sandboxing, and flexible deployment options ("Gartner recognized Codex’s strengths across agentic software development, enterprise governance, sandboxing, and flexible deployment options.").
Recent improvements to Codex include the integration of GPT-5.5, enhanced tool use, faster performance, and deeper support for enterprise workflows, enabling companies like Cisco to accelerate software development significantly ("Since Gartner's evaluation earlier this year, we have significantly improved Codex with the introduction of GPT‑5.5, as well as stronger tool use, faster performance, and deeper support for enterprise software development workflows.").
This recognition and the continued advancements in Codex demonstrate the growing maturity and enterprise readiness of agentic AI for software development, shifting focus towards safe and scalable deployment.
27 May 2026impact 75
Cisco integrates Codex for enterprise engineering
Cisco's deep integration of OpenAI's Codex into its enterprise engineering workflows demonstrates a significant advancement in agentic coding, transforming it into an AI engineering teammate.
Codex wrote over 95% of new AI features.
Defect resolution throughput increased by 10-15x using Codex CLI.
Over 1,500 engineering hours are saved monthly, partly due to a ~20% reduction in cross-repository build times.
Codex now reasons across large, interconnected repositories, works fluently in complex languages (C/C++), executes CLI-based compile-test-fix loops, and operates within enterprise governance frameworks.
Cisco is also leveraging Codex and GPT-5.5-Cyber through OpenAI's Daybreak initiative for cyber defense.
This case study provides strong evidence for the real-world, enterprise-scale impact and capabilities of agentic coding, pushing the frontier of AI as an active engineering teammate.
27 May 2026impact 65
Warp's Open Agentic Development and Oz platform
Warp introduces Open Agentic Development and its Oz orchestration platform, significantly advancing the practical application and scaling of AI agents in software development.
Warp's agents now co-create around 90% of the company’s internal pull requests, demonstrating a high level of integration and capability in real-world engineering workflows.
The new Open Agentic Development model outlines a process where humans define objectives and supervise outcomes, while agents handle planning, coding, testing, and opening pull requests.
Warp's Oz cloud orchestration platform manages persistent and parallelized agents across various environments, utilizing techniques like context compaction and persistent memory to ensure reliability in extended workflows.
Efficiency improvements are noted with GPT-5.5, which reduced token usage by 30% per agentic coding task compared to GPT-5.4, aiding in scaling long-running agent workflows.
This marks a significant step towards autonomous software engineering, with practical frameworks and platforms emerging to manage and scale agentic coding workflows in production environments.
27 May 2026impact 65
Codex-driven self-improving agent for tax preparation
Tax AI, a self-improving agent developed by OpenAI and Thrive Holdings, demonstrates significant advancements in automating complex tax preparation for accounting firms.
Tax AI processed 7,000 tax returns across Crete firms during its pilot season, saving practitioners a third of their time, drafting returns with up to 97% accuracy, and increasing throughput by 50%.
The system's correct field completion rate improved from 25% at launch to 86% within six weeks, driven by a self-improvement loop incorporating expert practitioner feedback, production traces, and a Codex-driven iteration process with tailored evaluations.
Codex investigates pipeline issues by inspecting traces, evals, repo, and skills, then implements targeted fixes, validates them, and proposes pull requests to close the improvement loop.
This case study provides a blueprint for building highly effective, self-improving agentic systems in complex real-world domains, showcasing the practical impact of advanced agentic coding capabilities.
27 May 2026impact 50
Local Speech Backend Deployment for Reachy Mini
privacy: Audio never leaves your network, the entire pipeline runs on hardware you control.; setup instructions: Detailed instructions for setting up the backend and connecting the robot are provided.
28 May 2026impact 65
Codex as full-lifecycle agent for 'agentic organization'
Endava has transformed into an "agentic organization" by deploying Codex as a full-lifecycle agent across its software development process.
Codex significantly reduced requirements analysis time from weeks to hours and exponentially improved output quality.
The system functions as a general desktop agent, assisting across the entire software lifecycle, including requirements, design, specifications, development, and operations.
It facilitates knowledge transfer and mentorship by codifying senior expertise and judgment, enabling junior engineers to produce senior-level outputs.
This demonstrates a significant expansion of agentic AI's application beyond just coding to encompass the entire software development lifecycle, highlighting its potential for organizational transformation and knowledge scaling.
2 Jun 2026impact 50
Launch of ChatGPT Sites
features: Hosting, Access controls, Storage, Database support; description: ChatGPT Sites enables the creation of tailored internal websites and lightweight apps for specific organizational tasks.; limitations: Cannot connect to live data
9 Jun 2026impact 60
Integration of Multimedia AI Components
source: Hugging Face Blog; description: A coding agent built a website featuring Paris monuments as 3D Gaussian splats.
17 Jun 2026impact 60
Integration of Strands Robots SDK with LeRobot
dataset format: The dataset format stays exactly as LeRobot wrote it; the agent loop is the glue.; workflow steps: The example agent in this post does four things: record new demonstrations in simulation, push the result to the Hub as a LeRobotDataset, run a policy in simulation against that same format, and deploy the same agent code to a physical robot with one keyword argument change.
18 Jun 2026impact 60
Benchmarking Tool for Coding Agents
token usage: agents used 1.3–1.8× (and up to 6×) fewer tokens.; agent usability: Library development must consider agent usability in design.
22 Jun 2026impact 60
Implementation of Local Models for Open-Source Contributions
focus: classification and triage processes; application: OpenClaw repository; local models: Gemma, Qwen
29 Jun 2026impact 65
Meta launches Pocket for AI-powered interactive app/game creation
Meta has launched Pocket, a new app enabling users to create interactive games and apps using AI prompts.
Meta's Pocket app enables users to create interactive apps and games with AI prompts.
Pocket was first launched on June 29, 2026, on the App Store and Google Play.
This marks a significant consumer-facing application of AI for software creation, potentially democratizing game and app development.
3 Jul 2026impact 45
Physical interface for AI-assisted shortcut creation
Project Mirage's Dune keypad introduces a physical interface for AI-assisted shortcut creation, leveraging Claude Desktop to generate context-sensitive automations.
The Dune is a three-key aluminum keypad that connects to a MacBook's USB-C port, drawing power directly and featuring context-sensitive buttons.
It integrates with Claude Desktop, allowing users to describe desired shortcuts in plain language for AI-assisted creation, which Claude then writes and assigns to a key.
A marketplace for user-generated skills is also offered, potentially making the hardware a front-end for a Claude-powered skills ecosystem.
This development extends agentic coding capabilities to a physical hardware interface, making AI-driven automation more accessible for general productivity tasks.
8 Jul 2026impact 60
Release of Nemotron V3 Data Atlas
pre training tokens: over 10 trillion; post training samples: millions
8 Jul 2026impact 60
SWE-bench Evaluation Improvements
human review agreement: 74; flawed tasks percentage: 30
9 Jul 2026impact 50
Launch of Muse Spark 1.1
features: advanced coding capabilities, bug detection, multimodal perception; model name: Muse Spark 1.1; free credits: $20 with new Meta Model API accounts; api availability: public API preview for US developers
9 Jul 2026impact 60
Launch of Muse Spark 1.1
pricing: {'input_tokens': 1.25, 'output_tokens': 4.25}; features: multi-step reasoning, complex workflows, agentic tasks; model name: Muse Spark 1.1
10 Jul 2026impact 55
Alibaba bans Claude Code usage
Alibaba has banned its employees from using Anthropic's Claude Code, citing its classification as high-risk software.
Alibaba will implement a ban on employees using Anthropic's programming tool Claude Code starting July 10, directing them to use Alibaba's internal Qoder tool instead.
This decision follows Anthropic's ongoing efforts to prevent Chinese companies and foreign entities owned by them from accessing its models, including closing loopholes for Chinese users.
This event highlights the growing complexities of geopolitical and corporate policy in the adoption and deployment of advanced AI tools, particularly in sensitive areas like coding.
Lineage
Descended from
This is one node. The map holds the whole field.
Watch Agentic Coding — and everything it connects to — grow as real news threads onto the map every day.