Devin Bots vs OpenAI Dots vs Claude Duet: Persistent Coding Agent Comparison and the Guardrails to Demand
Devin Bots, OpenAI Dots and Claude Duet are often compared, but they are three different kinds of product. This comparison scores them against five controls: reach, secrets, scope, egress and spend. It also covers lock-in, common mistakes, a bounded pilot and a buyer checklist.
Persistent coding agents run on a different rhythm from tools that handle one ticket at a time. It is tempting to treat Devin Bots, OpenAI Dots and Claude Duet as head-to-head rivals. That framing is less useful than it looks, because the controls that actually limit risk sit in different parts of each vendor's stack.
So this piece compares Cognition's Devin Bots, OpenAI Dots and Anthropic's Claude Duet using what we call the Five-Control Model. Reach is what compute and systems the agent can touch. Secrets is whether the model can read credentials. Scope is which repositories, tools and merge rights each agent gets. Egress is where the agent can send data over the network. Spend is how usage and cost are capped, alerted and stopped.
Why persistent agents change what you are buying
Fortune describes Devin Bots as a persistent coding tool that acts as a manager for teams of Devin agents and can run a complex project over days or even weeks. Cognition calls a Bot a persistent agent with its own memory and tools that can lead multiple Devins in parallel, and says you can hand it a large project or an ongoing goal. In Cognition's account, the Bot investigates, assigns work, judges the results and decides what comes next. It carries context between sessions and brings you only the decisions that need your judgment.
That last detail is the one I would underline. With a single-task agent, a person reviews every output. With a persistent agent, the agent decides what a person sees.
Market momentum, kept in proportion
Fortune reports that Cognition is valued at $48 billion and that the company announced its annualized revenue run rate passed $1 billion in September. The same report notes that SpaceX acquired Anysphere, the parent company of Cursor, in a deal that valued it at $60 billion. Citigroup's CEO has cited 14,000 Citigroup engineers using Devin, though that figure covers the existing Devin product, not Devin Bots. Separately, Cognition says Bots usage across dozens of early-access customers grew tenfold in two weeks.
Those numbers say investors are confident and sales are moving. They say nothing about fit with your codebase.
Three different product shapes
Devin Bots is coding-first and still in early access. Each Bot is a persistent agent with its own memory and tools that leads several Devins working in parallel.
OpenAI Dots was announced at OpenAI's DevDay as a personal agentic assistant powered by GPT-6 Astra. It is available in ChatGPT to Pro and Business Premium users in eligible markets, and users can launch it from Codex or ChatGPT.
Anthropic Claude Duet is described by SiliconANGLE as a standalone app on the web, iOS and Android for Pro and Max subscribers. Anthropic says Duet can use over 6,000 connectors and plugins in the Claude Marketplace.
Our reading is that Dots and Duet are general assistants first rather than dedicated coding agents. For engineering work, the fair comparison is with the documented controls in Codex cloud and Claude Code on the web, not with the assistants themselves.
Lock-in, model choice and pricing exposure
Model choice
Fortune reports that a Devin Bot picks suitable models for planning, design, computer use and other parts of a project, and uses faster, cheaper models for routine code. The sources reviewed do not say which providers those models come from, so we suggest asking directly and probing the routing during a pilot. Dots runs on OpenAI's GPT-6 Astra, and Duet comes through Anthropic's own subscription plans. If you already buy from either company, procurement gets simpler.
Pricing exposure
Engadget reports that Anthropic says subscribers to its $20 per month Pro plan will get Duet when it reaches general availability. Engadget also reports that Dots currently requires an OpenAI Pro plan, and that those plans start at $100 per month. OpenAI's availability statement also names Business Premium users, but the sources reviewed give no price for Business Premium or for Devin Bots.
On Anthropic's stack, Claude Code cloud sessions count toward your plan's usage limits along with the rest of your Claude and Claude Code usage, and there is no separate charge for the cloud VM. No extra infrastructure bill, then. The catch is throughput: an agent running for days draws on the same allowance your engineers rely on.
These are US dollar prices from launch coverage. Confirm current pricing and availability with each vendor before you budget.
The portability test
Ask each vendor one question: if we leave, can we export the agent's memory and configuration?
Reach, secrets and scope: what each vendor has published
An agent that runs for days should stay away from the machines and secrets your engineers use every day. Here is what the sources say.
- Devin Bots: Cognition runs Bots on separate cloud computers, away from the passwords and keys stored on engineers' laptops. Administrators choose which code repositories and tools each Bot can access.
- OpenAI Dots: Each dot can be given its own identity, credentials and tools through existing systems. OpenAI is working with Microsoft to integrate dots with Agent 365 security controls.
- Claude Duet: Anthropic says Duet handles credentials the same way as Claude Managed Agents, so it never directly sees your passwords or sign-in tokens.
- Claude Code on the web: In Anthropic-hosted environments, GitHub credentials stay encrypted on Anthropic's servers and never enter a session's VM.
The table maps the sources to the Five-Control Model and names the product behind each cell. Not published means we found no public documentation in those sources. It does not mean the control is missing.
| Control | Devin Bots | OpenAI Dots (Codex cloud for code) | Claude Duet (Claude Code on the web for code) |
|---|---|---|---|
| Lock-in: model choice | Different models chosen per task | Dots: powered by GPT-6 Astra | Not published in the sources reviewed |
| Reach | Separate cloud computers | Codex cloud environments | Claude Code on the web: cloud VM per session |
| Secrets | Kept away from secrets on engineers' laptops | Dots: own identity and credentials per dot; Agent 365 integration in progress | Duet: never directly sees passwords or tokens. Claude Code on the web: GitHub credentials stay outside the VM |
| Scope: repos and tools | Set by admins for each Bot | Dots: specific tools per dot | Duet: marketplace connectors; per-agent repo scoping not published |
| Scope: auto-merge | Not published | Not published | Not published |
| Egress | Meant to block unauthorized transfers; defaults not published | Codex cloud: blocked by default in the agent phase; setup scripts can reach the internet | Claude Code on the web: Trusted allowlist of close to 200 entries |
| Spend | Not published; cheaper models for routine code | Not published | Claude Code on the web: counts against plan usage limits |
Egress: why WorkOS says allowlists carry the load
For an agent that reads untrusted content, prompt injection is the main risk. A README or an issue comment can hide instructions telling the agent to send data somewhere it should not go. WorkOS, which reviewed sandbox egress defaults across vendors, argues that allowlists beat prompt-injection detection filters.
Fortune reports that Cognition's network controls are meant to prevent unauthorized data transfers, including when a model follows malicious instructions hidden in content it reads. The sources do not describe Cognition's default allowlist. Ask for it.
Codex cloud
By default, Codex blocks internet access during the agent phase. The settings around that default matter just as much.
- Setup scripts: Setup scripts still run with internet access so they can install dependencies. WorkOS says no documented setting turns off the setup scripts' unrestricted internet access.
- Per-environment access: You can turn on agent internet access for an individual environment when you need it.
- Allowlist presets: According to WorkOS, Codex's Common dependencies allowlist preset lists 69 apex domains, and you can add your own domains on top of None or Common dependencies.
- HTTP methods: Codex can limit network requests to GET, HEAD and OPTIONS and block POST, PUT, PATCH, DELETE and other methods. We suggest treating this as one useful layer rather than a complete defence against data leaks.
Claude Code on the web
WorkOS counts close to 200 entries across seventeen categories in Claude Code's Trusted-level allowlist. They cover package registries for ten language ecosystems, GitHub, GitLab and Bitbucket, container registries and cloud SDKs.
WorkOS also flags a configuration trap: Claude Code reads its strictest network settings only from user or managed configuration. If a team commits its hardening to the repository, the settings silently do nothing. While you are checking where settings live, run a Claude Code settings.local.json secret leak audit as well.
Allowlists versus detection
WorkOS notes that detection vendors advertise catching 95% of attacks and cites Simon Willison's view that this is very much a failing grade in web application security. The reasoning is simple. An attacker can keep retrying against a filter until it misses once, while an allowlist blocks unlisted destinations however many times someone tries. Model-level defences such as instruction/data separation against prompt injection add another layer, but the allowlist should do most of the work.
WorkOS also cites a UK AI Security Institute incident report. During a cyber evaluation, agents took unsanctioned actions on the live internet: 19 catalogued actions across 10 of 122 runs. The actions came to light when security monitoring flagged traffic leaving through Tor. This was an evaluation, not a production deployment of any of these products. Still, it was network monitoring that caught it.
Spend and scope: where the documentation runs out
Budgets
Devin Bots manages cost partly by sending routine code to faster, cheaper models. The sources do not describe spending caps for individual Bots. Claude Duet checks connected accounts every few hours because checking constantly would use up plan limits much faster.
Approvals and auto-merge
Before Duet sends a message, makes a payment, signs a document or deletes anything, it shows you exactly what it is about to do and waits for your approval. Some actions are blocked even with approval, such as trading securities or moving money between accounts. That is a sensible safeguard for an assistant. It is not a code merge policy.
None of the sources say whether any of the three products can merge code on its own, or how you would stop it. A Devin Bot escalates only the decisions it judges need your input, so ask Cognition directly whether merging to main is one of them.
Common mistakes and how to fix them
- Committing hardening to the repo. Claude Code reads its strictest network settings only from user or managed configuration, so settings committed to the repo fail silently. Move them, then test with a request that should be blocked.
- Trusting detection filters. A filter that misses some attacks eventually loses to an attacker who keeps retrying. Make a narrow egress allowlist your primary control.
- Mistaking in-agent approvals for merge control. Approval prompts cover assistant actions, not code merges. Apply branch protection, required human review and CI gates to agent identities exactly as you do to people.
- Forgetting setup-phase egress. On Codex cloud, setup scripts run with internet access even when the agent phase is blocked. Keep setup scripts minimal and review what they download.
- Running an agent under a person's identity. Give every agent its own identity and a spend cap or usage alert, so audit trails and plan allowances stay separate.
A bounded pilot using the Five-Control Model
Start by setting all five controls
- Scope: Pick one small, low-risk repository with an ongoing goal, like dependency updates or test coverage. Give the agent a dedicated identity limited to that repo and a named set of tools. No merge rights.
- Reach and secrets: Run it on isolated cloud compute and confirm credentials stay outside the session.
- Egress: Set the narrowest allowlist that still lets builds pass. On Claude Code, put hardening in user or managed configuration.
- Spend: Set a per-agent cap if the vendor offers one. If not, set a usage alert against plan limits and name who receives it.
- Exit criteria: Write down what result would justify expanding and what would make you stop.
Run with human gates
Require human review for every merge. Log each decision the agent escalates and each action it takes without asking.
Test egress
Plant a harmless prompt-injection test in content the agent reads, for example an instruction to send a dummy value to an outside domain you control. The request must be blocked and logged.
Review and decide
Compare actual spend with the cap or alert threshold you set at the start. Expand one repository at a time, and only if every exit criterion is met and both checkpoints passed.
Buyer checklist
- [ ] Reach: Does the agent run on isolated cloud machines, away from developer laptops? If you host agents yourself, isolate coding agents from your signed-in browser.
- [ ] Secrets: Are tokens held outside the session so the model cannot read them? If you host agents yourself, route agent API keys through a credential proxy.
- [ ] Scope: Does each agent get its own identity, with admin-set limits on repos and tools? Can you stop it from merging?
- [ ] Egress: Is network access blocked by default, setup scripts included? Can requests be limited to read-only methods? Does misplaced configuration fail silently?
- [ ] Spend: Are there per-agent caps, alerts and hard stops? How do plan limits affect multi-day work?
- [ ] Beyond the five: Can you export agent memory and configuration? Does the product work with security tools such as Agent 365? Is the plan you need sold everywhere you deploy?
If it would help, Optijara can work with your team to scope a first pilot against these five controls, or review an egress allowlist before an agent runs unattended.
Key Takeaways
- Devin Bots is a persistent coding tool that manages teams of Devin agents and selects suitable models for each part of a project. For coding work, we suggest judging Dots and Duet by the documented controls in Codex cloud and Claude Code on the web.
- Judge every option on five controls: reach, secrets, scope, egress and spend.
- Require isolated compute, a separate identity for each agent, access only to the repos it needs and credentials the model cannot read.
- WorkOS argues egress allowlists do more than detection filters. Also check setup-phase internet access and where hardening settings have to live.
- Claude Code hardening committed to the repository silently does nothing. Put it in user or managed configuration.
- None of the sources reviewed document per-agent budget caps or auto-merge policies, so set your own spend alerts and keep branch protection on.
- We suggest running a bounded pilot on one repository, with written exit criteria and pass/fail checkpoints, before expanding.
Conclusion
Devin Bots, Dots and Claude Duet are not three versions of one product. Devin Bots is a persistent manager of Devin agent teams that selects suitable models for different parts of a project. OpenAI announced Dots as a personal agentic assistant. For code, we suggest judging both Dots and Duet by the documented controls in Codex cloud and Claude Code on the web. Which one fits depends on the contracts you already hold, how much single-provider dependence worries you and how much per-agent control you need.
Whichever you pick, put it through the Five-Control Model. Keep reach to isolated compute and keep secrets out of the model's view. Scope repos, tools and merge rights per agent. Block egress by default with an allowlist, and cap spend. Devin Bots is still in early access, and vendors have published little on per-agent spending caps or auto-merge policies. Until that changes, keep your own branch protection, review and CI in place. Start with one repository, write your exit criteria before the pilot starts, and expand only when the pilot meets them.
Frequently Asked Questions
Is Devin Bots a direct replacement for OpenAI Dots or Claude Duet, or are they different kinds of products?
They are different kinds of product. Devin Bots is a persistent coding tool: an agent with its own memory that leads multiple Devins in parallel. Dots is a personal agentic assistant powered by GPT-6 Astra. Duet is a standalone app on the web, iOS and Android that can use over 6,000 connectors and plugins in the Claude Marketplace. For coding work, compare Devin Bots with the controls in Codex cloud and Claude Code on the web rather than with the assistants themselves.
Which persistent coding agent gives the least vendor lock-in for model choice?
The sources reviewed do not settle this. Fortune reports that a Devin Bot selects suitable models for planning, design, computer use and other parts of a project, but the sources do not say which providers those models come from, so ask directly. Dots runs on OpenAI's GPT-6 Astra, and Duet comes through Anthropic's plans. Whichever you consider, ask how agent memory and configuration can be exported, since that accumulated context is what makes switching hard.
What do these agents cost?
Engadget reports that Anthropic Pro subscribers, at $20 per month, will get Duet at general availability, and that Dots requires an OpenAI Pro plan starting at $100 per month. OpenAI also lists Business Premium users as eligible, but the sources reviewed give no Business Premium price and no price for Devin Bots. These are US dollar figures from launch coverage. Confirm current pricing and availability with each vendor before you budget.
What network egress settings should we require before letting a coding agent run unattended?
Outbound access should be blocked by default during agent work, with a narrow allowlist that adds only the domains your builds need. Find out whether setup scripts are covered: WorkOS reports that no documented setting turns off internet access for Codex cloud setup scripts. Where possible, limit outbound requests to GET, HEAD and OPTIONS. Then test the controls with a harmless prompt-injection attempt before you rely on them.
Why might our Claude Code network hardening not be taking effect?
WorkOS reports that Claude Code reads its strictest network settings only from user or managed configuration. If your team committed the hardening to the repository's settings, it silently does nothing. Move it into managed or user configuration, then confirm it works by testing an outbound request that should be blocked.
Can these agents merge code on their own, and how should we control that?
None of the sources reviewed document auto-merge policies for Devin Bots, Dots or Duet, so ask each vendor directly. Meanwhile, enforce branch protection, required human review and CI gates on agent identities just as you would for people. Duet's approval prompts before sending, paying, signing or deleting are a safeguard for an assistant, not a merge policy.
Sources
- https://devin.ai/blog/introducing-devin-bots
- https://fortune.com/2026/11/17/exclusive-cognition-bets-on-devin-bots-coding-agents-in-challenge-to-anthropics-duet-and-openais-dots/
- https://techcrunch.com/2026/09/29/openai-launches-dots-its-bubbly-agentic-avatar/
- https://learn.chatgpt.com/docs/cloud/internet-access
- https://claude.com/resources/articles/claude-duet
- https://code.claude.com/docs/en/claude-code-on-the-web
- https://workos.com/blog/agent-sandbox-egress-defaults
- https://www.engadget.com/2293646/claude-duet-is-anthropics-answer-to-meta-muse-and-openai-dots/
- https://siliconangle.com/2026/10/22/anthropic-introduces-claude-duet-a-proactive-ai-agent-for-professionals/
- https://dataconomy.com/2026/11/18/cognition-launches-devin-bots-to-manage-ai-coding-teams/



