Agentic harnesses are products and operating surfaces, not just model names. These buyer guides cover adoption, safety, pricing, and the setup decisions that comparison tables cannot answer alone.
ACE recommendation: Use Cursor (or VS Code + Cline) for continuous editor work; test Claude Code, OpenCode, or Muse Code for longer terminal objectives — many teams keep both.
ACE recommendation: Start with a small repository test set, keep branch protections intact, publish one AGENTS.md source of truth, restrict MCP to a short allowlist, and ship the review checklist with the tool choice.
ACE recommendation: Estimate cost per successful task (including retries), treat hybrid plans as seat-plus-metering, and keep a fallback provider for bursts, outages, and expensive loops.
A myth-versus-reality check for coding assistants whose pricing, editor surfaces, privacy paths, and autonomy are easy to collapse into one misleading label.
What actually matters
Claude Code is not Max-only — Pro, Team, and Enterprise paths exist; the API is a separate billing route from chat subscriptions.
Cursor Hobby is a limited entry plan, not an unlimited daily agent budget for teams.
Devin Desktop (local IDE, formerly Windsurf) and Devin Cloud (browser delegated agent) are different products with different setup boundaries.
ACE recommendation: Use the exact product surface, account or API path, endpoint, and review gate as the source of truth. Pilot on a disposable branch, check live terms before spend or sensitive code, and treat autonomy as a reason for stronger review rather than less.
ACE recommendation: Record the answers per product surface before approving a rollout. Price baseline seats plus realistic agent usage and a contingency in the AI stack calculator, pilot with an MCP and model allowlist, and require a human review gate for generated code and cloud changes.
Pick a local inference stack and wire it to an agent harness without the usual traps — wrong ports, tiny default context, bad quants, and mismatched API prefixes that make agents look broken when the model is fine.
What actually matters
Default ports differ by runtime: Ollama listens on 11434, LM Studio on 1234, llama.cpp server commonly on 8080 — point each harness at the correct host:port.
Agent workflows need 16k–32k context; Ollama’s small default can silently drop files unless you set OLLAMA_CONTEXT_LENGTH or Modelfile num_ctx.
For Aider with Ollama, use the ollama_chat/model prefix — not bare ollama/ — per official Aider docs.
ACE recommendation: Start with Ollama or MLX on a practice repo, set context before the first agent task, stay at Q4_K_M or better, match each harness’s API prefix, and keep a cloud escalation path for tasks local models stall on.
Roo Code’s discontinuation pushed many teams toward Kilo’s OpenCode-lineage stack. This guide covers Marketplace install, credential migration, project rules, and the VS Code vs JetBrains paths without assuming settings auto-port.
What actually matters
Install from the official Marketplace path (kilocode.Kilo-Code on VS Code); JetBrains plugin feature depth may differ from Roo’s old surfaces.
API keys and provider auth do not auto-port — expect full re-authentication on every provider.
Kilo CLI tokens often live under ~/.local/share/kilo/, not ~/.config/kilo/ — check before debugging missing credentials.
ACE recommendation: Install Kilo from the official Marketplace path, re-authenticate every provider, consolidate rules into AGENTS.md, pilot on a sanitized repo, and read Kilo’s platform docs before assuming Roo parity.
Many teams keep Cursor for the editor loop and Claude Code for terminal delegation — but dual-running without boundaries creates duplicate edits, conflicting plans, and review gaps.
What actually matters
Assign surface ownership: Cursor for continuous editor work; Claude Code for multi-file terminal objectives and git-heavy automation.
Parallel work needs git worktrees or separate branches — never two agents on the same branch without coordination.
One AGENTS.md keeps instructions consistent when both tools read the same repository.
ACE recommendation: Treat Cursor as the default editor loop and Claude Code as the escalation path for longer terminal work; use worktrees, one AGENTS.md, and the same review checklist — never two agents on one branch without coordination.
Agent harnesses read different instruction files by convention. Without one portable source of truth, switching tools or running two agents scatters conflicting teaching across the repository.
What actually matters
Each harness reads different files by convention — AGENTS.md, CLAUDE.md, .cursor/rules, extension configs — without coordination they diverge.
Conflicting instructions (test command, style, forbidden paths) produce confident wrong patches when two agents run on one repo.
Project memory should be reviewed like code when the stack or policies change.
ACE recommendation: Publish one AGENTS.md (or equivalent) in the repo root, link or copy into each harness’s native format, and treat instruction drift as a bug — not an inevitable side effect of multi-tool teams.
A hub for the OSS and open-client cluster — Cline, OpenCode, Aider, Kilo, Goose, Continue (frozen hosted path), Amp, and Warp Agent CLI — and when bring-your-own-key or local inference is cheaper than a bundled subscription.
What actually matters
Client type shapes fit: full IDE agents (Cline/Kilo), terminal git workflow (Aider), OpenCode-lineage TUI (OpenCode/Kilo CLI), terminal CLIs (Goose, Warp Agent).
Continue’s OSS extension works for local/BYOK assistive chat, but its hosted roadmap is frozen — new rollouts should evaluate Cline/Kilo.
BYOK plus local Ollama can beat a bundled seat when usage is predictable and hardware is already paid for.
ACE recommendation: Shortlist two OSS clients against one paid seat on the same repo test set; BYOK wins when usage is predictable and local hardware is already paid for — paid seats win when onboarding speed and unified billing matter more than marginal token savings.
Amp and Claude Code both target terminal-first agent work, but install paths, billing (Amp credits/Orbs vs Claude Pro/Team pools), and review habits differ enough that a one-week pilot should decide — not a feature checklist.
What actually matters
Install and auth paths differ: Amp’s ampcode.com web sign-in then `amp` vs Claude Code’s Pro/Team/API routes — confirm you can authenticate in one session.
First-week cost includes subscription, credits/Orbs, API overage, and retry burn — not just the headline plan price.
Both need the same review gate: plan, tool log, diff, tests before merge.
ACE recommendation: Run the same three repository tasks on both for five days, track cost and review time, then pick one primary terminal agent — dual-running without worktrees wastes money and creates duplicate edits.
MCP servers extend agents with filesystem, shell, and network tools. Treat each install like granting operator access — allowlist servers, scope folders, fail closed, and log calls; ignore files do not block MCP.
What actually matters
Each MCP server install is equivalent to granting filesystem, shell, or network access — not a harmless plugin.
Org allowlists and approval workflows must exist before developers install servers ad hoc.
Folder and network scope per server must match the repo’s data classification.
ACE recommendation: Ship a short MCP allowlist with the rollout guide, require security review before any server reaches production repos, and audit tool logs weekly during the first pilot month.
Autonomy is a reason for stronger review, not weaker gates. This checklist ties to the rollout guide: inspect the plan, tool log, diff, tests, and rollback path before any agent-generated patch merges.
What actually matters
Read the agent’s plan before approving tool calls — especially shell, git, and MCP actions.
The diff must match the stated objective without drive-by refactors, deleted tests, or secret leakage.
Tests must run on the same branch the agent touched, or document a explicit reason why not.
ACE recommendation: Pair every harness rollout with this checklist and the coding-assistant-rollout-guide — measure accepted patch rate and review minutes, not vanity autonomy metrics.
