Coding agents compared: the table that starts with workflow

Most coding-agent comparisons put products from four categories into one feature grid and call every checkmark equal. This page starts with the work surface, then compares control, isolation, cost, context, automation, and review in the order you will experience them.

By the Continuum team. We build a workbench that runs Claude Code, Codex, and their peers, so the model rates quoted here are the ones our own cost analytics ship with.

The short version

Use an editor agent when you steer continuously, a terminal agent when you delegate bounded repository work, a cloud agent when the task should run away from your machine, an open-source harness when provider control or self-hosting matters, and a workbench when several sessions create an orchestration problem. Cursor and Copilot lead the editor class; Claude Code and Codex lead terminal delegation; Devin leads managed asynchronous work; OpenCode, OpenHands, and Aider cover different open-source needs; Conductor and Continuum manage parallel agents rather than replacing them.

What you need to know
  • Choose the work surface first: editor, terminal, cloud queue, open-source platform, or workbench.
  • A model benchmark cannot tell you about permissions, branch hygiene, review time, or quota failure.
  • Cursor and Copilot optimize steering; Claude Code and Codex optimize delegation.
  • OpenCode, OpenHands, and Aider are three different open-source product shapes.
  • Devin sells managed distance from the loop; that is useful only when the task survives distance.
  • A workbench earns its place when session state and review queues become the bottleneck.

The comparison table

Read the second column before any feature. Product shape predicts how the tool feels, what it can contain, and when the human returns. A terminal agent can be better at a whole ticket while an editor agent is better for the next ten minutes. Neither result contradicts the other.

Public product and pricing material checked in August 2026. Free software can still incur model, hosting, or hardware cost.

ToolShapeCode runsCost entryBest fit
CursorAI-first editor + CLILocal or Cursor cloud agentHobby free; Pro $20/moContinuous steering and predictive editing
GitHub CopilotEditor extension + CLI + cloud agentLocal editor, CLI, or GitHub cloudFree allowance; Pro $10/moExisting editors and GitHub-centered teams
Claude CodeTerminal + desktop/web surfacesLocal, remote, or Anthropic cloud surfaceIncluded by eligible Claude plans or APIBounded repository delegation
OpenAI CodexTerminal + app + cloudLocal sandbox or cloudIncluded by ChatGPT plan level or APIChatGPT users and sandboxed terminal work
OpenCodeOpen-source TUI, desktop, editor, serverLocal or your chosen provider pathMIT software; model cost separateProvider choice and local models
AiderOpen-source terminal harnessLocal processApache 2.0 software; model cost separateSmall harness, automatic commits, model control
OpenHandsOpen-source autonomous platformDocker sandbox, hosted, or self-hostedMIT core; compute and model cost separateSelf-hosted autonomous workflows
DevinManaged autonomous service + desktop/CLIManaged environmentsFree entry; paid individual and team plansAsynchronous ticket delegation
ConductorMac multi-agent workbenchLocal worktrees or cloud sandboxes by planFree local tier; paid cloud tiersDense parallel work on macOS
ContinuumMulti-provider, multi-device workbenchEnrolled local or remote hostsFree app; hosted inference optionalSeveral agents, accounts, devices, gauges, and one ledger

The table intentionally avoids a single quality winner. Cursor's best feature is present while you type. Devin's best feature is that you do not have to be present. Aider's best feature is source-level control and a git trail. Continuum's best feature is seeing several independent agent sessions as one operating system. Converting those into one number would hide the choice rather than clarify it.

The first split: steer, delegate, or assign

Every switcher query becomes easier after one verb. You either steer code as it changes, delegate a bounded task while remaining close to the repository, or assign work to run elsewhere and return later. The interfaces follow from those verbs.

Working modeHuman loopBest product shapeTypical tools
SteerSeconds: read, accept, reject, redirectEditor agentCursor, GitHub Copilot
DelegateMinutes: describe, inspect plan, review diffTerminal agentClaude Code, Codex, OpenCode, Aider
AssignHours: submit ticket, check status, review resultManaged cloud agentDevin, OpenHands Cloud, vendor cloud agents
OrchestrateAcross many sessions: route, approve, queue reviewMulti-agent workbenchConductor, Continuum

Problems begin when a team buys one shape and expects another. An editor agent feels needy if the goal was to assign a ticket and leave. A hosted agent feels opaque if the engineer wants to watch every tool call. A terminal agent feels fragmented when six branches are running and nobody can see which one is blocked. Product dissatisfaction is often interface mismatch mislabeled as model quality.

Permissions and isolation compared

A coding agent is a process acting with credentials on files and instructions it did not write. The meaningful comparison is not whether a settings page contains the word security. It is which boundary remains when the model makes a bad decision or reads a hostile instruction.

Control shapes, not a certification claim. Verify exact defaults in the linked documentation before deployment.

Tool classPrimary controlUseful boundaryCommon mistake
Cursor / Copilot editorPer-action approvals and organization policyEditor workspace plus cloud-agent isolation where usedAssuming a familiar editor makes shell access harmless
Claude CodePermission modes, allow/deny rules, hooksFilesystem and command policy; worktree isolation for parallel sessionsUsing bypass mode on the normal laptop because prompts seem trustworthy
CodexApproval policy plus sandbox modeOS sandbox using Seatbelt on macOS and Landlock on LinuxGranting network or broad writes to an unattended run without a need
OpenCode / AiderConfigurable permissions or process permissionsYour OS account, container, VM, or local model boundaryConfusing open source with sandboxed execution
OpenHandsDocker sandbox and mount policyContainer boundary around agent commandsMounting a broad host path read-write and treating the container as isolation
Managed cloud agentVendor environment, repo token, organization policyShort-lived remote workspace if configured correctlyGiving the installation access to every repository and production secret
Worktree workbenchOne checkout and branch per sessionCollision and rollback boundary, not a secret boundaryBelieving git isolation also isolates Keychain, SSH agent, or network credentials

Worktrees prevent two agents from overwriting the same checkout and make each diff attributable. They do not isolate processes, credentials, databases, ports, or the network. Containers and OS sandboxes narrow process reach. They do not fix an overpowered token mounted inside the boundary. Repository-scoped credentials, short lifetimes, and an explicit network policy still matter.

The safest default for unfamiliar work is read-only planning, followed by workspace-scoped writes, followed by human review before any external side effect. Full autonomy belongs in an environment that can be discarded and whose credentials cannot deploy, merge, delete data, or enumerate unrelated repositories.

Cost and limits are different product decisions

A $0 label can mean free software, free inference, a small quota, a local model, or a workbench that expects you to bring a paid provider. Those are different economics. Compare the complete route from prompt to model and back.

Cost shapeExamplesWhat happens under heavy use
Seat with usage windowClaude Code on Claude plans, Codex on ChatGPT plansThe window closes or credits extend it, depending on plan
Seat with monthly allowance or creditsCursor, GitHub CopilotPremium work slows, stops, or moves to metered usage by product
Pay per tokenAider, OpenCode, SDK use with API keysNothing stops unless you set a provider budget
Managed subscription and quotaDevin and cloud-agent productsIncluded work ends, then the product stops or sells more usage
Free app plus your providersContinuumThe app stays free; each selected provider enforces its own bill or quota
Local modelOpenCode, Aider, OpenHands with local inferenceNo token invoice, but hardware, power, latency, and maintenance remain

The practical comparison is interruption versus exposure. A hard quota protects the budget and can strand a task. Metered overflow preserves the task and can produce an invoice. Local inference protects against both vendor meters and can be slower or weaker on the repository you care about. None is inherently best. Pick the failure you can schedule around.

For a team, price review time separately from model spend. A $10 seat that produces twice the rejected code of a $20 seat is not cheaper. Track completed tasks, review minutes, rework, and model cost together for at least two weeks. A provider dashboard can show tokens. It cannot tell you whether the diff was accepted.

Context, rules, and extension surfaces

Agent quality depends on what the harness loads before and during the task. Every serious tool now supports project instructions, but the files and extension points differ. Cross-tool teams should keep durable rules in the simplest shared form and add vendor files only for behavior that has no portable equivalent.

NeedPortable routeTool-specific depth
Repository conventionsAGENTS.md where supportedCLAUDE.md, Cursor rules, tool settings
On-demand workflowsPlain scripts and documented commandsClaude skills, Cursor rules, agent-specific commands
External systemsMCP where the client supports the same serverVendor connectors and managed integrations
Deterministic policyCI checks, branch rules, OS permissionsHooks, permission callbacks, organization policy
Programmatic agentHeadless CLI with structured outputClaude Agent SDK, OpenHands SDK, OpenCode server and SDK paths

The portability rule is simple: place facts close to the code, place enforcement outside the model, and keep vendor prompts thin. Build commands, architecture boundaries, and test gates belong in version control. Merge protection belongs in GitHub or the forge. Secret scope belongs in the credential issuer. A model instruction may repeat those facts, but it should not be the only place they exist.

Extension count is not a pure advantage. Hooks, MCP servers, subagents, rules, skills, and custom commands all consume context or add failure modes. Load what a task needs. A small repository with one build command does not benefit from a forty-page agent constitution.

Editor agents: Cursor versus GitHub Copilot

Cursor and Copilot overlap more each year, but the adoption decision remains different. Cursor is the editor. Copilot enters the editor you already use. Cursor can tune the full interaction between index, completion, chat, and agent. Copilot can meet a mixed organization in VS Code, Visual Studio, JetBrains, and GitHub without asking everyone to move.

DecisionCursorGitHub Copilot
Best reasonPredictive editing and one coherent AI-first surfaceBroad editor coverage and GitHub-native procurement
Free startHobby with limited usage and no expiryMonthly free completion and chat allowance
Paid entry$20/mo Pro$10/mo Pro
CLI / headlessCursor agent CLICopilot CLI and GitHub automation surfaces
Cloud workCursor cloud agentsGitHub coding agent and pull-request workflow
Switching costEditor migrationExtension and account rollout

Choose Cursor when predictive editing is the product you are buying and the team accepts an editor standard. Choose Copilot when editor diversity, GitHub policy, and low-friction rollout dominate. If autonomous ticket work is the main requirement, compare their agent surfaces with Claude Code, Codex, or Devin separately instead of letting autocomplete decide the purchase.

Terminal agents: Claude Code, Codex, OpenCode, and Aider

Terminal agents share the same visible loop: read files, call tools, edit, test, report. They differ in the machinery around the loop. Claude Code offers the deepest first-party extension system. Codex offers explicit sandbox and approval controls plus a strong path for ChatGPT subscribers. OpenCode offers open code and broad providers across several surfaces. Aider offers the smallest harness and the strongest default git trail.

PriorityBest starting pointReason
Deep project automation and agent extensionsClaude CodeHooks, skills, subagents, MCP, headless mode, and Agent SDK
OS sandbox and an existing ChatGPT billCodexExplicit sandbox modes and plan inclusion
Provider choice, local models, source accessOpenCodeMIT client and broad models.dev provider route
Automatic commits and minimal machineryAiderEach edit becomes a reviewable git boundary by default
A neutral comparisonThe same task in two worktreesRepository evidence beats a generic winner

Do not trial terminal agents on a one-file function. Give them a task that requires locating code, respecting a constraint, changing a test, and running the real gate. The point of the harness is the loop. A prompt that never exercises tools measures the model and misses the product.

Autonomous and workbench products

Devin and OpenHands occupy the autonomous side. Conductor and Continuum occupy the orchestration side. The former pair do work in environments. The latter pair organize sessions run by other agents. They can appear similar because all show tasks, terminals, and diffs, but the ownership boundary differs.

QuestionAutonomous platform answerWorkbench answer
Who is the agent?The product or its configured runtimeClaude Code, Codex, Cursor, OpenCode, or another hosted CLI
Who owns compute?Vendor cloud or your self-hosted platformYour enrolled machine, plus optional cloud paths by product
Primary valueWork continues away from the developerMany sessions remain visible and isolated
Review unitReturned task, branch, or pull requestOne live session and worktree at a time
Choose whenAsynchronous assignment is the goalParallel local or remote agents are already the reality

Conductor is the focused Mac answer for parallel worktrees and a dense board. Continuum spans more providers and devices, adds per-account quota gauges and a local cost ledger, and keeps the app free with your own subscriptions. This site builds Continuum, so verify those claims on the product and pricing pages rather than treating this comparison as independent review.

A workbench should not increase the number of active agents by default. Its first job is to make existing work legible: ownership, plan state, diff, tests, quota, and review queue. Add concurrency only after those signals are reliable.

A switcher test that survives the next model release

Model leads move. A useful migration test measures the product around the model and can be repeated after any release. Use one normal repository task, one risky read-only task, and one failure. The failure is important because agents look most alike when everything works.

01

Normal task

Choose a two-hour issue with a known test. Record elapsed time, human prompts, files touched, checks run, and review minutes.

02

Risky task

Ask for a plan involving auth, data, infrastructure, or a destructive migration. Confirm the tool can stay read-only and makes the boundary visible.

03

Failure task

Use a deliberately failing test or unavailable dependency. Observe whether the agent diagnoses, loops, broadens scope, spends without a ceiling, or stops cleanly.

04

Parallel task

If multi-agent work matters, run two independent tasks in separate worktrees and measure the review queue rather than the number of running sessions.

05

Exit test

Export or inspect the diff, session record, configuration, and branch without the product. Switching cost is part of the score.

MetricWhy it matters
Accepted lines after reviewOutput volume without acceptance is noise
Human interventions per taskShows steering burden more honestly than elapsed runtime
Review minutesThe scarce resource once generation is cheap
Rework within seven daysCatches plausible code that did not hold
Model and software cost per accepted taskMakes free, seat, and token routes comparable
Time to recover from failureTests whether git, session history, and isolation are operational

Keep the tool that reduces total time from issue to accepted change. That answer may be a pair: an editor agent for steering and a terminal agent for delegation, or a native agent plus a workbench once parallel sessions become routine. Buying one product for ideological simplicity is not simpler if the work still has two shapes.

Questions people ask

Start with work surface: steering in an editor, delegating in a terminal, assigning to a cloud environment, owning an open-source platform, or orchestrating several sessions. Then compare task completion, review time, isolation, cost behavior, and switching cost on the same repository task.

An AI editor keeps the human in a seconds-long loop around completion and edits. A coding agent can read files, run tools, change several files, execute tests, and return a diff after minutes. Modern products overlap, but the dominant interaction still matters.

Claude Code offers a deeper first-party harness for many workflows. Codex offers explicit sandbox controls and can be the better value when ChatGPT already covers it. The same bounded task in clean worktrees is a better answer than a generic benchmark.

Cursor is the stronger choice when predictive editing and one AI-first editor justify migration. Copilot is the stronger choice when a team wants AI in its existing editors with GitHub-native policy and a lower paid entry.

OpenCode is a polished open-source coding agent with several surfaces and many providers. OpenHands is a self-hostable autonomous platform with Docker sandboxes and SDK paths. Aider is a smaller terminal harness with broad model choice and automatic git commits.

No. Worktrees isolate checkout and branch state, which prevents collisions and improves rollback. They do not isolate credentials, processes, network access, ports, or external services. Use an OS sandbox or container plus scoped credentials for those boundaries.

Sometimes only the software is free. Check model inference, cloud compute, hardware, and the review time the tool creates. A free harness using a paid frontier API is metered; a local model avoids the token bill but consumes hardware and maintenance.

When two or more sessions create a visibility problem: branch ownership, plans waiting for approval, diffs waiting for review, quota state, or spend spread across providers. If one terminal session is legible, stay native.

Sources

Every figure above was read from these pages on August 2026. Vendors reprice without notice; if you find a stale number, tell us.

  1. Claude Code documentation
  2. OpenAI Codex sandboxing
  3. Cursor pricing
  4. GitHub Copilot plans
  5. OpenCode documentation
  6. Aider documentation
  7. OpenHands documentation
  8. Devin documentation
  9. Conductor documentation
  10. Continuum pricing
Try it

Compare agents.
Keep one workbench.

Continuum runs several official coding agents under your own plans, gives each session an isolated worktree, and puts plan, diff, PR, quota, and cost state in one place. The app itself is free.

free app · your subscriptions · local-first