6. Platform Operations

Platform Operations is where you manage the big picture — autonomous code generation, human oversight, strategic goals, and project organization. These features give your autonomous agents the structure they need to operate safely and effectively at scale.

Forge: Autonomous Feature Implementation

Forge (/forge) is Sutra's autonomous coding engine. Describe a feature in plain language, and Forge will plan it, write the code, run tests, and open a Pull Request — all without manual intervention.

Placeholder: Forge Terminal

How Forge Works

  1. Request: Submit a feature description and target GitHub repository (e.g., "Add dark mode toggle to settings page in owner/repo").
  2. Planning: Forge clones the repository, analyzes the codebase, and generates a step-by-step implementation plan using an LLM.
  3. Plan Approval: Review and approve the plan before any code is written (or enable auto-approval for trusted workflows).
  4. Coding: A LangGraph-powered coding agent implements each step — reading files, writing code, and making changes across the codebase.
  5. Testing: Forge automatically detects and runs your project's test suite (pytest, npm test, etc.) and reports results.
  6. Pull Request: Once tests pass, Forge creates a feature branch, commits the changes, and opens a PR on GitHub.

Forge supports feedback rounds — if the plan isn't right, provide feedback and Forge will revise it. The entire lifecycle is tracked: queued → planning → awaiting_plan_approval → coding → testing → pr_created → completed. You can also trigger Forge requests directly from Telegram using the /forge command.

Tip: Forge works best when the feature request is specific and well-scoped. "Add a dark mode toggle to the settings page" produces dramatically better results than "improve the UI." You can configure the default LLM provider and model for Forge in System Settings.

Approvals (Human-in-the-Loop)

An autonomous organization without oversight is a risk. The Approvals queue (/approvals) acts as a critical safety boundary.

Placeholder: Approvals Queue

Managing High-Risk Actions

Certain tools (e.g., send_email, run_shell_command, commit_and_push) can be flagged as "Requires Approval." 1. Interception: When an agent attempts to use a flagged tool, Sutra intercepts the call and pauses the agent's execution. 2. Review: The action appears in the Approvals queue, showing the exact payload the agent intended to send. 3. Intervention: - Approve: The agent resumes execution immediately. - Modify: You can manually edit the payload (e.g., fix a typo in the email draft) and then approve. - Reject: You provide written feedback (e.g., "Your tone was too informal"). This feedback is sent back to the agent, who will attempt to correct its behavior and submit a new request.

The approval system is fully configurable per tool, per agent, and per risk level. You might auto-approve a research agent's web searches but require manual review for any agent sending external emails. This granularity lets you move fast where it's safe while maintaining tight control where it matters.

Tip: Use the "Modify" option generously in the early days. When you edit an agent's proposed action and approve the corrected version, the feedback loop helps you identify which system prompts need refinement — if you're constantly fixing the same type of mistake, the prompt needs updating.

Goals & OKRs

Goals (/goals) shift the paradigm from micro-managing tasks to defining high-level objectives.

Placeholder: Goals Tracker

Setting Objectives

Instead of creating a detailed Jira board, you define a Goal such as: "Migrate the frontend from React 17 to React 18." - Manager Agents: You assign this goal to a Manager Agent. - Autonomous Breakdown: The Manager Agent analyzes the codebase, creates the necessary sub-tasks, and assigns them to available Worker Agents. - Tracking: The Goals dashboard provides a visual progress bar, updating automatically as the worker agents complete their assigned tasks.

Goals are intentionally high-level — they describe the what, not the how. The Manager Agent figures out the how: breaking down the objective into tasks, assigning them to the right workers, and coordinating dependencies. This mirrors how effective human managers operate — set clear objectives and trust the team to execute.

Best Practice: Write goals with clear success criteria. "Improve the landing page" is vague. "Redesign the landing page hero section with a new headline, updated screenshot, and mobile-responsive layout" gives the Manager Agent enough specificity to plan effectively.

Projects

Projects (/projects) serve as isolated workspaces to keep your organization's efforts focused.

Placeholder: Projects Dashboard

Organizing by Project

If your organization is working on "Marketing Q3 Launch" and "Backend Refactor", creating separate projects ensures: - Focused Context: Agents assigned to a specific project will only search the Knowledge Base documents explicitly attached to that project, reducing hallucinations. - Dedicated Teams: You can assign specific agents exclusively to a project, ensuring they aren't distracted by global Tasks or unrelated Discussions. - Budget Isolation: Each project can have its own token budget, preventing one team's heavy usage from starving another project of resources. - Knowledge Scoping: Attach specific Knowledge Base collections to each project. Agents working on "Marketing Q3 Launch" will only search marketing documents, reducing noise and improving answer accuracy.

When a project wraps up, you can archive it to keep the dashboard clean. Archived projects retain their full history — conversations, tasks, workflow runs — and can be restored at any time for reference or audit purposes.