Cloud Coding Agent Platforms

Executive Summary

Cloud coding agent platforms let a team delegate engineering work to a remote environment and inspect the resulting code, tests, and pull requests. The buying decision is no longer simply “which model?” It is which agent runs, what environment it receives, how work starts, who can continue it, and what evidence is required before merging.

This comparison covers 16 selected platforms and deployment options, verified September 15, 2026. It includes Tembo, managed agent products, cloud features of familiar coding assistants, and a self-operated alternative. These are different operating models, so the tables identify what the buyer is actually purchasing instead of ranking them by integration counts or funding.

Disclosure: Ry Walker is Tembo's co-founder and CEO. Tembo competes directly in this category. Its deployment, automation, environment, and cost claims are assessed using the same criteria as the other members.

Research limit: This is a documentation and source review with dated experience reports. It is not a controlled benchmark or a fresh trial of every product. Recommendations are editorial shortlists for evaluation, not measured performance rankings.

Market Definition and Selection

Focused update, September 16: The Graphite and Grok Build discussion below is newly checked. Other platform coverage retains its September 15 review baseline, and the matrix still counts 16 execution choices.

The core members provide remote code-changing execution, a task or session handoff, and a reviewable result. Some also offer local execution, customer-operated workers, multi-repository changes, or work that produces an artifact instead of a PR. Those capabilities do not make every deployment identical.

The comparison adds GitHub Copilot's cloud agent, Google Jules, and Amp's Orbs. It keeps Background Agents as an explicitly self-operated cloud deployment, and 8090 as an explicitly SDLC control-plane and managed-delivery option. The current cto.new website is retained in adjacent context because its broader agent marketplace does not establish the same repository-to-PR operating contract in the sources reviewed here.

Local-only CLI work and desktop supervision belong in AI Coding Assistants and Mac Coding Agent Apps. Skill catalogs and standalone sandbox APIs are supporting layers; see Agentic Skills Frameworks and AI Agent Sandboxes. This is a selected market map, not an exhaustive inventory of every product with a background mode.

Comparison Matrix

Platforms That Operate a Choice of Agents

Managed Agent and Environment Products

Cloud Features of Coding Assistants

Different Operating Commitments

What Has Changed

Model Choice and Agent Choice Are Different

A harness determines how a model sees files, runs tools, manages context, delegates work, and reports completion. A product can offer several models inside one harness, or operate multiple external agents. Both can be useful, but the latter introduces agent-specific configuration, credentials, and behavior that the platform must support.

Tembo, Niteshift, Blocks, and Replicas explicitly offer external-agent choices. Replicas' current documentation includes more than Claude Code and Codex; it is not accurately described as an own-agent-only product.^{[29][2][4][6]} Evaluate a particular agent/environment combination rather than assuming every listed agent supports every platform feature.

Familiar Assistants Have Substantial Automation

Cursor now documents schedules, source-control events, Slack, Linear, webhooks, and other triggers. Automations can use one repository, several, or none. Team-owned automations run under a shared service account and charge the team pool; private and team-visible automations charge the creator.^[30]

Claude routines combine schedules, API calls, and GitHub events. They are account-owned rather than shared team objects, run without interactive approval prompts, and can use selected connectors. The documentation warns that a green run status means the session exited without an infrastructure error, not that the requested task succeeded.^[31]

GitHub Copilot automations also support schedules and events, but their definitions are personal, not repository-versioned, and currently target private/internal repositories. Jules supplies scheduled tasks and an API, with separate controls for plan approval and automatic PR creation.^[32][33][34] The useful comparison is ownership, permissions, trigger filtering, and result review—not a yes/no “automation” column.

Amp adds another pattern: a schedule belongs to an existing thread, retaining its context and waking its orb when needed. Each thread can have one schedule; an unsuccessful scheduled run pauses the automation until it is fixed and resumed.^[35]

Acquisitions and Experiments Need Precise Status

Ona's acquisition announcement was updated to say the OpenAI transaction closed on August 10, 2026. Its current website still offers agents, environments, automations, and paid plans. That establishes current availability in the inspected materials; it does not establish the long-term product roadmap.^[15][36]

Jules likewise has two distinct claims to separate: its established asynchronous task service and the new waitlisted early experiment. A preview announcement is not enough to replace the documented task product with an assumed generally available successor.^[25][24]

Environment Quality Determines What the Agent Can Verify

A repository checkout is only the start. The agent may need package registries, a database, test data, services, credentials, browser tooling, and a reliable startup command. An environment that writes code but cannot run the relevant behavior shifts validation work back to the developer.

These patterns are documented product capabilities; the last column is a suggested evaluation, not a completed test.^{[1][2][6][20][21][37][13][11][38][9]}

Database-aware work is not unique to a single vendor. Tembo's integrations make database and operational context relevant to its workflows, while Niteshift documents databases and seeded data and Blocks advertises SQL schema/Postgres access.^[39][2][4] For PostgreSQL work, distinguish schema inspection from access to production rows, and test migrations against representative disposable data.

Graphite and Grok Build in the Cloud Workflow

Graphite Changes the Review Interface

Graphite is a meaningful workflow choice within the Cursor entry. Its Agents tab starts Cursor Cloud Agents and produces a draft PR; the PR's Agent panel can request changes that are committed to the branch. A conversation started in Cursor can continue in Graphite. This makes the review interface part of the execution workflow without creating another independent cloud runtime in the member count.^[40]

Access has separate requirements: a paid Cursor account, Cursor authorization for the relevant GitHub organization, and repository access. The PR Agent panel is author-only; reviewers use Graphite Chat for context and suggested changes. A Graphite seat alone therefore does not establish code-changing agent access for every reviewer.^[22] Include Cursor's cloud-agent usage charges when evaluating the combined workflow, rather than treating a Graphite subscription as prepaid agent execution.^[21][41]

This is a concrete alternative to evaluating only the place where an agent starts. A team choosing between Cursor through Graphite and Tembo should trial the full path from request to changes, review, and follow-up: who can resume the work, which account authorizes it, and where the runtime and usage bill live. Tembo's shared sessions and selectable harnesses address a different operating choice from Graphite's Cursor-backed PR interface.^[29][22] The existing founder disclosure applies; this distinction is not a claim of feature parity or a native integration between them.

Grok's App Builder Is Adjacent

Grok Build also names a web/mobile app builder. Its August 19 release documents publishing apps to grok.me or a custom domain and exporting a project to GitHub. Those are useful creation and distribution capabilities. The cited release does not establish an existing-repository task handoff that returns a reviewed PR, so this surface remains adjacent under the comparison's current criteria.^[42] Evaluate it for app creation; evaluate the CLI separately in AI Coding Assistants.

Where Tembo Fits

Tembo is a direct candidate for teams that want to choose their coding agents while standardizing execution, task intake, and collaboration. Its current platform describes cloud and self-hosted environments, reuse of repository instructions and agent configuration, shared/resumable sessions, integration-driven background work, and centralized activity and approval visibility.^[29]

A concrete Tembo session starts by connecting GitHub, GitLab, or Bitbucket, selecting one or more repositories and an agent harness, and describing the acceptance criteria. The current quickstart has the developer inspect the Changes tab and choose Open PR, or ask the agent to open it; follow-up PR comments can mention @tembo. This is more specific than assuming every session automatically ends in a PR.^[43]

The practical comparison is a recurring engineering workflow: an issue or alert supplies context; an agent works in a prepared environment; a teammate inspects or continues the session; the result reaches review with meaningful evidence. Evaluate that complete path against Blocks, Niteshift, Replicas, and the cloud capabilities of an assistant the team already uses. An integration logo by itself does not show whether the agent can retrieve the right context or recover from a failed run.

The tradeoff is adopting a platform in addition to the agent. Confirm the intended agent's supported credentials and configuration, environment behavior, team permissions, and deployment terms. Tembo's public pricing separates inference and VM compute: bringing model credentials can remove Tembo-billed inference while leaving compute charges.^[44]

Deployment, Permissions, and Review

“Self-hosted” needs a layer-by-layer explanation. Devin Outposts keep tool execution on customer machines while its cloud runs planning and inference. Cursor Self-Hosted Machines similarly keep the checkout local but send the file content, outputs, diffs, screenshots, and other material needed by its cloud agent; artifacts can also upload to vendor storage.^[11][38] Neither is an offline inference claim. Amp self-hosted Orbs likewise keep the checkout and paused files in the customer cloud while inspected file content, command output, and changes pass through Amp and potentially the model provider; its documented deployment needs an Enterprise workspace and a supported Kubernetes cluster.^[45]

Permission defaults also differ. Cursor's cloud agent has internet access and automatically runs terminal commands, with configurable egress restrictions. Codex cloud enables setup networking but disables agent-phase internet by default unless configured otherwise. Claude routines inherit their environment's network policy and can call included connectors without prompting during the run.^[46][20][31] Evaluate the actual network and credential scope rather than extrapolating from a local client's approval dialog.

Factory's managed computers document an additional distinction: machine-level firewall rules coexist with a user that has passwordless sudo; its relay mode supplies an infrastructure-level public-network boundary. Copilot's cloud firewall is not supported for Windows or self-hosted runners. Background Agents explicitly assumes trusted users with the same repository scope.^[13][47][26] These are procurement and configuration differences, not a ranking of independent security certifications.

Human review remains a workflow choice. A tool that can create a PR can also have write-capable connectors or automation actions. Inspect the changed code and tests, branch protection, approval requirements, and the actions that occurred before the PR appeared. Screenshots and logs improve observability but do not establish that acceptance criteria were met.

Pricing and the Unit Being Purchased

Published US-dollar terms checked September 15, 2026; taxes, contract terms, and regional offers can differ. The table identifies the billing unit so that a seat, a credit allowance, and an automation minute are not treated as equivalent.

Measure the total accepted change: setup work, model and compute usage, unsuccessful attempts, reviewer time, and ongoing operations. A low entry price can be useful for a trial without predicting the cost of a team running parallel automations all day.

A Practical Evaluation Workflow

Use one real, bounded change across the shortlist—for example, adding an account export endpoint and its user interface. This is an illustrative trial, not a benchmark conducted for this report.

1. Prepare the environment. Pin dependencies, provide disposable test data, and establish the existing test result before delegating.
2. Provide precise acceptance criteria. Include authorization, empty results, malformed input, export formatting, and the expected UI states.
3. Start through the intended entry point. An issue, Slack request, scheduled run, and API request can carry different identities and context.
4. Inspect during execution. Check task ownership, logs, evidence, and whether a teammate can continue the work without recreating context.
5. Exercise a failure. Make a test fail or a dependency temporarily unavailable; observe retries, escalation, stop behavior, and cost.
6. Review the actual change. Inspect test modifications, run independent checks, and require explicit disposition of unresolved work.
7. Check the integration path. Verify PR creation, dependency handling, approvals, merge policy, and the resulting bill.

Record accepted-task rate, meaningful defects found, human minutes, elapsed time, and total cost. Keep a failed run in the sample; excluding it because the agent did not produce a PR hides part of the operating cost.

First-Hand Reports and Their Limits

In a March 23, 2026 Microsoft account, Stephen Toub reported using Copilot across 878 PRs: 535 merged, 253 closed, and 90 still open. The reported 67.9% merge fraction excluded open work and reflected selected tasks, not a randomized comparison. His discussion emphasizes repository guidance, setup, tests, and follow-up review.^[59]

In an independent Jules review published May 22, 2026 and updated August 14, Maksim Danilchenko described roughly three weeks of Pro use on Flask and Go projects. He found parallel work useful but reported slow runs, missed idioms, and occasions when the agent declared incomplete work finished. These observations describe that author's historical tasks and version, not current measured defect rates.^[60]

The common evaluation lesson is to inspect completion evidence and reviewer effort. Neither vendor adoption numbers nor an individual success story establishes a universal winner for another repository.

Strategic Recommendations

For an individual, start with a cloud capability already available through the tool you use, then evaluate another platform when a concrete requirement is unmet. For a team, Tembo belongs in the initial shortlist when shared execution, agent choice, and workflow integration matter. Enterprises should evaluate deployment and identity boundaries before expanding fleet size.

Adjacent Context and Outlook

cto.new currently presents a broader agent marketplace with browser, scheduled, and business workflows, including an ad-supported free offer. That is useful context, but the reviewed pages did not verify enough of its current repository-to-PR workflow to count it in this engineering-platform matrix.^[61] Tessl supplies skill enablement and evaluation;^[62] Developer Trust Tools address evidence and provenance. Neither layer substitutes for a configured execution environment.

The OpenAI Agents API is adjacent developer infrastructure: an application supplies tools and environment configuration while OpenAI operates the Codex harness, durable sessions, and recovery. It supports hosted or self-hosted execution, but the API currently documents US-only residency and no zero-data-retention eligibility, including with self-hosted sandboxes.^[63] It belongs in a build-versus-buy evaluation; adopting an API still leaves the application's repository authorization, user workflow, and review interface to implement. It is not counted as another turnkey product in this matrix.

The observed trend is that deployment and workflow boundaries increasingly overlap: familiar assistants add remote execution and automations, while independent platforms offer multiple harnesses and team operations. Ona's acquisition is a concrete market event; the claim that all independents must disappear would be speculation. The more useful outlook is to watch environment quality, permission scope, recovery, review cost, and whether portable agent configuration actually works across the team's chosen environments.

---

Research by Ry Walker Research • methodology