Switch from Augment Code to Continuum safely

Decide whether to switch

A full switch makes sense when Augment's distinctive context and editor capabilities are underused and the larger problem is coordinating several coding-agent sessions. Continuum gives each supported agent a separate git worktree and keeps its chat, plan, diff, pull request, terminal, and artifacts in one place. It can expose sessions through Mac, web, iPhone, and narrow Watch controls. Windows and Linux desktop clients are beta.

Keep Augment when Context Engine materially improves cross-repository work, JetBrains integration is essential, inline assistance is heavily used, Slack codebase questions are embedded in team practice, or Cosmos provides required organization workflows. Continuum has no direct substitute for those capabilities. It coordinates other agents and branches. The underlying provider still supplies model execution, repository reasoning, and its own tools.

A partial move is often the strongest outcome. Keep Augment in VS Code or JetBrains for context-intensive work. Use Continuum for parallel tasks assigned to supported CLI agents. Where policy and compatibility permit, expose Augment Context Engine MCP to the underlying agent. This gives every component a clear job and avoids pretending that a workbench contains a code index.

Inventory before changing anything

Create a written inventory while the current system still works. Record every repository and branch scope, VS Code and JetBrains workspace, Augment rule file, .augment-guidelines, agent definition, MCP server, context connector, Slack channel, remote environment, secret, source-control integration, issue workflow, review requirement, model setting, and usage owner. Include informal habits that are absent from configuration.

For each item, assign one disposition: preserve, translate, replace, or retire. Name an owner and acceptance test. "MCP moved" is too vague. A useful entry says which agent loads the server, where configuration lives, which identity it uses, which repositories it can read, what success query proves it works, and how the connection fails when access is revoked.

capability: cross-repository architecture retrieval
source: Augment Context Engine
destination: retained MCP through chosen agent
owner: developer-platform
test: 20 known-answer questions, zero restricted-repo leaks
rollback: Augment IDE remains installed through release N+1
status: pilot

Export only human-readable and permitted data. Preserve repository instructions, approved prompts, saved runbooks, audit material required by policy, and monthly usage baselines. Augment's semantic index, ranking state, conversation history, and remote execution state do not become Continuum data. Assume vendor-specific histories will stay behind unless the product provides an explicit supported export.

Choose the replacement agent

Continuum supports Claude Code, Codex, Cursor, Gemini, Grok, and OpenCode. The provider choice determines model behavior, command tools, approvals, sandboxing, context limits, instruction precedence, MCP support, usage billing, and enterprise terms. Continuum does not normalize those differences. Select the provider before designing the rest of the migration.

Use representative work. Include one small bug, one monorepo change across packages, one database or wire-contract migration, one browser-visible issue, one task that reads history, and one failure that requires interpreting tests. Run each with the same acceptance criteria and repository permissions. Track prompts, context corrections, elapsed time, usage, review comments, repairs, and accepted results.

If no replacement agent matches Augment on cross-repository retrieval, choose among three honest outcomes. Retain Context Engine through MCP where supported. Keep Augment for the affected repositories. Accept the weaker retrieval path and compensate with explicit repository instructions, search tools, architecture documents, and smaller tasks. Record the tradeoff rather than hiding it inside a broad migration status.

Check commercial and security fit at the provider layer. Continuum is free, though the selected agents use their existing subscriptions, API accounts, or included plans. Confirm data use, retention, identity, audit, regional needs, model availability, rate limits, and purchasing ownership for every provider that will enter production.

Translate rules, context, and tools

Move durable repository guidance into reviewed files whenever possible. Augment supports workspace guidance and files such as .augment-guidelines, and its documentation recognizes common repository instruction files. The destination may use AGENTS.md, CLAUDE.md, provider-specific rules, skills, MCP configuration, or other formats. Read the selected agent's precedence rules before copying text.

Remove instructions that describe Augment UI, tool names, or environment assumptions. Preserve verifiable constraints: build commands, test targets, formatting, generated-file policy, ownership, prohibited paths, security boundaries, and definition of done. Split broad files by directory when the provider supports scoped instructions. Conflicting global and nested rules create silent drift.

Context Engine requires a separate decision. If retained, configure its MCP endpoint in the underlying agent and test authentication from a new worktree. Verify repository permission trimming, citations, freshness, unavailable-service behavior, and whether the worktree branch is represented. If retired, document which local search, repository map, language server, code search, or architecture source replaces each high-value query.

Recreate tools with new credentials. Git hosting, issue trackers, package registries, cloud consoles, browsers, databases, and deployment systems should receive scoped identities tied to the new workflow. Avoid copying a broad token from Augment configuration into several agents. Test denial paths and revocation before enabling write access.

Map category workflows

Define state transitions for every category: queued, owned, running, waiting for input, ready for review, accepted, merged, failed, and abandoned. A branch should have one owner and one current status. Do not launch the same ticket in Augment and Continuum as a race unless the duplication is an explicit evaluation with separate branches and a named comparison process.

Set concurrency by review capacity. Start with two active implementation sessions per reviewer. Increase only when review queue age, conflict rate, rejection rate, and post-merge defects stay within target. Parallel agents can create output faster than a team can understand it. Worktrees preserve files; they do not preserve attention.

Pilot and cut over

The acceptance gate should include all affected surfaces. Verify plan review, code edits, focused tests, full build where appropriate, diff review, pull-request creation, terminal access, mobile control, usage visibility, credential revocation, worktree cleanup, and integration after concurrent branches merge. A successful chat response proves almost none of this.

Keep implementation and rollout status separate. A repository may be technically ready while security review, purchasing, mobile pairing, or provider access remains incomplete. Report each state directly. This prevents a pilot success from being interpreted as organization-wide migration completion.

Rollback and cleanup

Rollback should be possible without reconstructing developer state. Keep Augment extensions and CLI configuration installed through the overlap period. Preserve untouched baseline branches, source-controlled instructions, environment bootstrap scripts, and a list of required connectors. Avoid deleting Augment organization data until the destination survives a release and audit requirements are satisfied.

Define rollback triggers before the pilot: restricted context leakage, unreviewed write access, repeated missed dependencies, accepted-change cost above the baseline, worktree corruption, review queue age beyond target, provider availability below target, or a critical workflow with no replacement. Give one person authority to stop the rollout.

At final cleanup, revoke Augment tokens and service accounts, remove Slack and source-host integrations, close remote environments, export billing records, document any retained Context Engine service, and remove obsolete local rules only after the replacement is confirmed. Then audit the chosen agents for stale credentials and duplicated MCP entries.

Git worktrees need routine hygiene. Merge or close completed branches, preserve work that remains under review, remove abandoned worktrees through normal git operations, prune stale metadata, and verify the primary checkout is clean. Never use broad deletion commands against unresolved paths. A worktree is recoverable when its branch and commit are still known.

What Continuum adds and leaves out

Continuum adds a shared operating surface around supported agents. Sessions can be started on a Mac or enrolled host, isolated by worktree, reviewed through plan and diff panes, inspected through terminals and artifacts, and controlled from web or iPhone. Quota gauges are available where providers expose them, and local Claude and Codex activity can be summarized by provider, repository, and day.

Continuum leaves editor choice and model service outside the app. It has no inline code completion, embedded text editor, proprietary semantic index, managed model bundle, Augment provider, Slack agent, or automatic import of Augment rules and history. Those boundaries keep the workbench focused and make the migration more explicit. They also mean a team may need another product for context retrieval or in-editor assistance.

Questions people ask

Keep reading