Analyst(s): Mitch Ashley
Publication Date: February 5, 2026
From mid-2025 through early 2026, software platform vendors introduced new approaches to agent-driven development that are forming along two parallel paths. One path emphasizes multi-agent execution, where coordination and parallelism are primary. The other emphasizes intent-first structuring, where specifications and constraints shape how agents act. These paths are not mutually exclusive. OpenAI’s Codex app establishes a concrete baseline for multi-agent execution, providing the market with a reference point for how intent-first structure and governance can be applied as agent work moves into production environments.
What is Covered in this Article:
- OpenAI’s Codex app establishes a baseline for multi-agent software development.
- Two parallel paths are emerging: multi-agent execution and intent-first structuring of agent behavior.
- IDE-centric tools are evolving toward multi-agent execution through different entry points and priorities.
- All approaches face the same constraint: enterprise-grade governance for parallel agent execution remains immature.
The News: In early February 2026, OpenAI announced its Codex app, a macOS desktop application designed to support AI-assisted software development using multiple agents. The app provides an interface for interacting with AI agents that perform software-related tasks, including code generation, modification, testing, and documentation across projects.
The Codex app introduces several capabilities intended to support parallel agent work. These include Skills, which define reusable agent capabilities; Automations, which allow agents to run tasks on a scheduled basis or in the background; and worktrees, which enable agents to operate in isolated environments while working on the same codebase. Agents can run for extended periods and request user approval when broader system access is required.
The application integrates with GitHub for repository access and pull request workflows and connects to the broader ChatGPT account environment. OpenAI states that Codex runs in a sandboxed environment by default, with restricted permissions, and supports configurable access controls for executing commands or modifying files.
According to OpenAI, the Codex app is powered by its Codex model family and is intended to support development tasks across multiple stages of the software lifecycle. The company positions the app as a standalone desktop experience rather than an extension of an existing integrated development environment.
Agent-Driven Development – Two Paths, One Future
Analyst Take — Two Parallel Paths Toward Agent-Driven Development: Agent-driven development is no longer theoretical. What is unfolding across the market are two parallel paths toward the same outcome: AI agents performing execution across the software lifecycle under human direction.
The first path is multi-agent execution. In this model, developers coordinate multiple agents working in parallel, assign tasks directly, and supervise outcomes as work progresses. Anthropic’s Claude Code and OpenAI’s Codex app exemplify this approach. Both treat agent coordination, parallelism, and long-running execution as first-class workflow elements, with structure and constraints introduced as execution unfolds.
The second path emphasizes intent-first (or requirements-first) structuring of agent work. Here, explicit requirements, design artifacts, and constraints shape how agents act before execution begins. Platforms such as Amazon Web Services’ Kiro, IBM’s Project Bob, and planning-oriented modes within Google’s Antigravity initiative emphasize intent capture as the control surface for agent execution. Agents still perform the work, but within boundaries defined upfront.
These paths are complementary rather than competitive. Both assume agents execute meaningful work. They differ primarily in sequencing: whether coordination and execution come first, or whether intent and structure come first. Increasingly, both approaches coexist within the same platforms and workflows as vendors blend coordination, execution, and upfront constraint in preparation for production use.
Multi-Agent Execution Enters Through Different Entry Points
Multi-agent execution is not emerging from a single tool or workflow. Vendors are entering the multi-agent path from different starting points, emphasizing different priorities while converging on parallel agent execution as a core capability.
GitHub’s Agent HQ represents a governance-first entry into multi-agent execution. It is designed as a centralized hub for running and managing multiple agents within a single project, with a focus on guardrails, identity, approvals, and sandboxed execution. By anchoring agent orchestration in DevOps and security workflows, GitHub treats agents as long-lived actors whose actions must be governed, auditable, and constrained from the outset.
Anysphere’s Cursor represents a developer-speed-first entry into multi-agent execution. Cursor supports running multiple agents in parallel on isolated workspaces, often using separate branches or worktrees, with developers reviewing and merging agent-produced diffs. This places Cursor closer to Codex-style multi-agent execution than to IDE augmentation, even though the IDE remains the primary interaction surface.
Windsurf takes a different approach. It is best described as a single-agent-centric, agentic IDE optimized for deep context, large codebases, and long-running sessions with a powerful primary agent. While multiple agent interactions are possible across separate contexts, Windsurf does not currently emphasize first-class parallel agent orchestration as Codex, Cursor, or Agent HQ does.
These approaches illustrate how multi-agent development is forming through varied sequencing rather than a single canonical model.
Why Codex Establishes a Multi-Agent Baseline For OpenAI
OpenAI’s Codex app matters because it establishes a clear baseline for multi-agent execution. It makes coordination, parallelism, and extended agent runtime a primary workflow rather than an experimental add-on. That baseline gives enterprises and competing vendors a concrete reference for evaluating what is required next.
Intent-first approaches can layer structure, policy, and compliance on top of this execution model. IDE-centric tools can evolve toward it. Codex does not define the end state of agent-driven development, but it does provide a starting point for understanding how multi-agent systems behave in practice.
The Shared Constraint: Governance
Across multi-agent execution and intent-first structuring, the same constraint applies. Parallel, persistent agent execution introduces coordination risk without a corresponding governance infrastructure.
Observable agent behavior, runtime policy enforcement, and provenance tracking remain incomplete across platforms. A common failure scenario illustrates the issue: one agent modifies infrastructure configuration, another updates tests based on a prior state, and a third generates documentation assuming success. No individual agent is “wrong,” yet the system becomes inconsistent and difficult to audit or recover.
This is why the future of agent-driven development will not be determined solely by execution sequencing. It will be determined by how effectively governance is integrated into agent execution itself.
What to Watch:
- Convergence signals: Where multi-agent execution platforms add intent capture and where intent-first platforms support more dynamic execution.
- Governance integration: When enforceable policy controls, evidence generation, and audit trails become native to agent workflows.
- Enterprise adoption patterns: How regulated industries sequence intent and execution differently from commercial teams.
- Augmentation limits: The point at which IDE-centric approaches fail under coordination complexity.
- Hybrid workflows: Whether platforms support mixing multi-agent execution and intent-first structuring within the same development lifecycle.
Read OpenAI’s Codex app announcement for full details.
Disclosure: Futurum is a research and advisory firm that engages or has engaged in research, analysis, and advisory services with many technology companies, including those mentioned in this article. The author does not hold any equity positions with any company mentioned in this article.
Analysis and opinions expressed herein are specific to the analyst individually and data and other information that might have been provided for validation, not those of Futurum as a whole.
Other insights from Futurum:
AI Reaches 97% of Software Development Organizations
100% AI-Generated Code: Can You Code Like Boris?
Dynatrace Perform 2026: Is Observability The New Agent OS?
Harness Incident Agent: Is DevOps Now The AI Engineers of Software Delivery?
Author Information
Mitch Ashley is VP and Practice Lead for the CIO & Technology Buyers and Software Lifecycle Engineering practices at The Futurum Group. A multi-time CIO and CTO with 30+ years leading technical organizations, Mitch built and operated production systems spanning cybersecurity for the U.S. Department of Defense, PKI services for the broadband and 5G industries, SaaS platforms, large-scale telecom and banking systems, and a national broadband network. His work with AI began early, developing expert systems that diagnosed and repaired complex mainframe environments. That operator foundation grounds his analysis in operational consequence, covering the technology buyer's world of software engineering, cybersecurity, DevOps, cloud, and AI.
