Choose an editor workflow before choosing a model
Continue provides open source extensions for VS Code and JetBrains, with Chat, Plan, Agent, Edit and autocomplete workflows described in its documentation. The important choice is what access a particular task needs. Chat has no tools; Plan uses read-only tools; Agent can make changes and run commands. The selected model or provider must support the required tool use. This is a documentation-led assessment checked on 6 November 2026, not a record of an installed extension or a completed coding benchmark.
Give it a change you can fully review
Use a disposable repository with a small function that parses a return-window value. Define the contract first: accept a non-negative whole number followed by the word days; reject a missing value, a negative number or extra trailing text. Prepare examples such as 30 days, 0 days, an empty string, -1 days and 30 days later. The desired deliverable is a minimal patch, tests for the stated cases, and a plain explanation of any behaviour the specification leaves open. Do not ask the assistant to redesign the surrounding application.
Separate understanding, changes and verification
- Inspect in Plan. Open the implementation and its existing tests, then ask Continue to identify the current behaviour, relevant files and proposed change. Its Plan documentation describes a read-only mode for this stage. Check the named files yourself; a plausible explanation is not proof that the right code was read.
- Choose one editing route. Use Edit for a selected section or Agent for a task that needs coordinated changes. Request the smallest patch that implements the agreed contract. Keep dependency installation, unrelated formatting and broad refactoring outside the exercise.
- Review tool permission. The Agent quickstart describes permission prompts and configurable tool policies. Inspect those settings before assuming every command will stop for approval. Allow only the actions needed in this disposable project. A rejected command should remain rejected rather than being replaced by an equivalent unreviewed route.
- Verify outside the conversation. Run the project's existing test command yourself, examine the actual diff and compare it with the five fixtures. Record whether the tests were executed, their result and which files changed. An assistant saying that tests pass is not a test receipt. Check that malformed input is rejected for the intended reason, rather than simply causing an unrelated exception.
Model configuration is part of the experiment
Continue documents separate model roles including chat, edit, apply, autocomplete, embedding and reranking. Record the provider and role used for the exercise; changing the role configuration can change both data flow and behaviour. Use the current config.yaml reference when adapting configuration. The documentation marks older config.json and some legacy context-provider material as deprecated, so an old tutorial may describe a different setup.
When it is a useful fit
Consider Continue when your team wants configurable AI assistance inside an existing editor and is willing to own model selection and code review. If the real requirement is unattended organisation-wide automation, first assess the separately documented CLI and operating controls rather than assuming an editor extension supplies them. The repository's core license is Apache 2.0; model, service and integration terms remain separate. A local editor also does not imply local inference. Keep the patch, fixtures and redacted configuration as the evidence from an eventual trial; no such trial or productivity result is claimed here.