Use the local workspace

Run agentagon in any local agent project. A GitHub remote is optional. The React workspace connects to a detached local service, so closing the browser does not cancel managed work.

On first use, choose Open local folder for an existing directory or Clone repository only when you need a new checkout. The folder does not need a GitHub remote, and code-only assessment also works outside Git. Detect agents scans the folder and makes valid detections available immediately. It preserves edits and excluded agents, needs no coding backend, and does not run application code. The optional Analyze code and traces disclosure starts a deeper assessment.

The main navigation follows the engineer's recurring jobs:

• Overview lets you choose an agent and shows work that needs input.
• Agents shows active and excluded agents. Open an agent directly; use Edit details, Exclude or Restore when needed.
• Production shows monitor coverage, observations, deployments, and evidence gaps.
• Activity shows tasks that need input, are running, or have finished.
• Settings configures the project, coding backend, connections, execution, and data boundaries.

Goals, issues, workflows, connections, and lessons remain available from the agent or action where they matter. They are not separate destinations that must be visited before doing useful work.

Fix a supplied failure

Choose All issues from Overview or an agent workspace, or use Fix a problem from an agent. Paste or upload supported trace JSON/JSONL, select an imported snapshot, choose a configured provider trace, or describe the failure directly. A saved goal is optional.

Pasted, uploaded, and provider traces use a review step. Choose Check trace first. Agentagon shows the detected format, trace identities, usable spans, incomplete coverage, redaction, and any blockers without retaining evidence. When the preview is usable, choose Import and investigate or Import and prepare fix to create the immutable project snapshot and continue. Editing the input invalidates the preview and requires another check.

Open a retained issue to review its evidence and lifecycle. You can assign an active agent, save reviewed expected behavior, mark the issue Not actionable with a reason, or reopen it. These decisions preserve the original evidence and use issue revisions to prevent one client from overwriting another.

Before work starts, Agentagon prepares the complete intent and displays ownership, source identity, verification prerequisites, expected outputs, and known limits. Submission rechecks the same revisions. Missing ownership or expected behavior becomes an explicit question instead of an unsupported repair.

A completed Fix or Optimize task opens a full result page. It separates the task's recommendation from your decision, compares the current version and verified candidates, shows the exact source diff and independent review, retains failed attempts, and lists limitations. Select this change or Keep current version creates a durable decision. Prepare local delivery creates reviewable local artifacts; it does not merge, push, publish, or deploy.

Improve and observe

Open an agent and choose Improve task success, Make tool use reliable, Make it faster, Reduce cost, or Write a goal. The goal page contains optional Details and one Go action. Go authorizes a bounded run to prepare and accept measurements, prepare the evaluator, measure current behavior, and search for independently verified improvements. Compatible evidence can be reused. Existing required checks remain protected.

The page shows what is happening and offers Pause, Continue and Stop according to the saved run state. When a workflow needs an answer, its question opens beside the goal. Answering lets the run continue. A missing source, credential or execution prerequisite shows the input that needs attention. Closing the page does not cancel the run; an interrupted service requires explicit continuation.

Runner and limits and Evaluation details are optional disclosures. They expose execution choices, cases, scoring and the evaluator editor without turning internal workflow stages into user chores. The default run allowance is at most 30 active minutes and 30 reserved evaluation trials, subject to configured limits. Values, directions, guard outcomes, sample identity and limits stay attached to results. Go does not select, publish, merge or deploy a result.

Monitoring is enabled explicitly for one active agent and environment. Production keeps collection state, comparison evidence, deployment declarations, and tested revisions separate. A verified code change does not become production recovery automatically. Missing samples, mixed releases, partial coverage, and stale collection remain visible limitations.

Tasks survive dashboard or MCP disconnection. Service interruption leaves work interrupted and requires an explicit resume. Reads do not start work or reconnect providers. Mutations require the exact local origin and session token.

Production monitoring · Memory groups · MCP · Workflow behavior