Coding-agent Behaviors

Coding-agent Behaviors are patterns Adaline finds across agent task runs. They are useful when you want to understand how a coding agent works over time: where it verifies changes, gets stuck, loops, skips tests, recovers from errors, or follows a healthy edit-and-check path. Use this page for logs coming from OpenCode or similar coding agents. For the general Behavior catalog workflow, see Understanding Behaviors.

​

Before Behaviors appear

Coding-agent projects need enough structured evidence for Adaline to understand the task path, not only the final answer. In addition to normal log traces and spans, send a stable task or run identifier, agent metadata, phase-like spans, outcomes, and safe summaries of important tool or command results. Some coding-agent projects may also require project enablement or a specific coding-agent tag/metadata convention before the coding-agent views are populated. If you are setting this up for the first time, contact support@adaline.ai or use the metadata your Adaline team provides. At a minimum, useful coding-agent logs should include:

• One trace or run boundary for the coding task.
• A stable session, task, or run ID when the task spans multiple traces.
• Agent name or type, workflow, repository or project category, language, framework, and environment when safe to send.
• Spans for planning, search, file reads, edits, tool calls, command runs, tests, recovery, and final handoff.
• Outcome/status, such as success, failed, partial, blocked, or needs review.

​

Read the coding-agent catalog

The catalog works like the general Behaviors view, but the rows are about coding-agent task patterns. A row might describe skipped verification, edit churn, runtime blockers, inconsistent summaries, repeated exploration, or healthy task completion. Read the row name, short summary, role, issue tag, workflow tag, represented evidence, and related objects. The row tells you what Adaline is seeing repeatedly; it is not enough by itself to change an agent prompt, tool policy, repo indexing, or command wrapper. Open the Behavior to read the triage and examples. Saved views are helpful for recurring reviews, such as issue-only coding-agent review, build-flow review, exploration-flow review, or verification-quality review.

​

Inspect triage findings

Triage findings explain why the pattern was flagged and what Adaline thinks happened. For coding agents, a triage result may call out the root process gap: for example, the agent applied a fix but did not run a reproducer, treated code inspection as verification, retried the same command without learning, or summarized a result that did not match the trace. Read the diagnosis next to the suggested fix and evidence links. Evidence links should take you back to exact spans, commands, model calls, or tool results. When available, use Open session to inspect the larger task journey. Good triage should help you decide whether the fix belongs in:

• The coding-agent prompt or workflow instructions.
• Verification policy, tests, command wrappers, or stopping criteria.
• Tooling, repository search, indexing, retrieval, or environment setup.
• Evaluation criteria or dataset coverage for future releases.

​

Use Sessions for task runs

The Sessions tab is useful when you want to start from task runs instead of a Behavior. Each row summarizes a coding-agent journey with its intent, outcome, trace count, and span count. Use this view to find representative successful, failed, partial, or unusually large runs before opening the full journey. The trace and span counts matter because coding-agent work is rarely a single call. A short task may only have a few spans; a hard investigation can include many tool calls, file reads, edits, retries, and verification steps. Open the session when the row suggests enough evidence to explain the behavior.

​

Review trajectories and phases

Member trajectories are representative task runs that contributed to the Behavior. They help you see whether the pattern is consistent across tasks or caused by one noisy example. You can open them from a Behavior detail page, or start from Traces and switch to the Sessions tab when you want to inspect coding-agent journeys directly. An opened trajectory shows a narrative, key actions, issues observed, phase count, turn count, and completion status. Use that summary before going span-by-span. When you open a trajectory or session, read it as a journey: Phases are especially useful for coding agents because the final answer can look acceptable while the path was fragile. A run that passed may still skip meaningful verification; a failed run may still provide useful examples for datasets, evaluators, or workflow instructions.

​

Common coding-agent patterns

Your actual labels depend on the evidence in the project. Use the labels as entry points, then rely on triage and trajectories for decisions.

​

Choose the next action