Current AI Agentic Setups: Developer Workflows and Tooling Trends

The Shift Toward Consolidated AI Workflows

Modern developer workflows are moving away from complex, multi-agent orchestration platforms and toward consolidated, high-context interactions. Rather than managing a "swarm" of specialized agents, many developers are using a single, high-capability model—most notably Claude 3.5 Opus—to act as a central orchestrator. This central agent manages sub-tasks, delegates to other sessions, and maintains the context required to manage complex software projects.

Subscription and Model Preferences

Most developers maintain a mix of subscription-based services and pay-as-you-go (PAYG) APIs. Claude Pro and ChatGPT Plus remain the dominant consumer subscriptions, often supplemented by PAYG services like OpenRouter for model experimentation or specific tasks that flagship models refuse to handle.

  • Claude: Frequently cited as the primary driver for coding and design tasks, often utilized via the Claude Desktop app or integrated into IDEs.
  • Codex: Remains a staple for backend-heavy tasks and reliable code generation.
  • Local Models: A subset of power users are increasingly relying on local inference using hardware like Apple Silicon (M1/M5 Max) with MLX or Ollama. This is driven by a desire for privacy, lower latency, and the ability to process asynchronous tasks without subscription overhead.

Coding Harnesses and Interfaces

Developers are increasingly opinionated about their coding harnesses, often choosing tools that prioritize "tight loops" of interaction over automated, hands-off agent behavior.

  • Preferred Tools: Claude Code, Aider, and OpenCode are commonly mentioned as primary interfaces.
  • The Human-in-the-Loop Preference: Many developers express skepticism toward "agentic" platforms that promise to replace human input. Instead, they prefer to steer the AI conversationally, using the agent to generate code or debug while retaining control over architecture and decision-making.
  • Customization: There is a strong trend toward building custom CLI harnesses or using terminal multiplexers (like Zellij) to manage multiple agent sessions simultaneously, allowing developers to load-balance tokens and maintain separate contexts for frontend and backend work.

The Decline of Complex Orchestration

While platforms like OpenClaw, Hermes, and various swarm tools exist, they have seen limited adoption among experienced developers. Common criticisms include:

  • Fragility: Complex orchestration often falls apart when tasked with real-world, non-trivial work.
  • Latency and Toil: Setting up and tuning these platforms is frequently described as "emailing your AI," introducing unnecessary friction compared to direct, conversational interaction.
  • Over-Engineering: Many users find that current models are capable enough to handle sub-tasking internally if prompted correctly, rendering external orchestration layers redundant.

Security and Privacy Concerns

As AI agents gain access to local files and APIs, security has become a primary concern. Developers are adopting several defensive strategies:

  • Sandboxing: Using tools like nono.sh to restrict file and network access for AI agents.
  • Credential Isolation: Creating unique, restricted API keys for specific AI services to ensure that a compromised agent cannot access sensitive production environments.
  • Local Inference: Running models locally remains the gold standard for processing sensitive data, such as personal emails or proprietary internal documentation.

Sources