Short answer: This week’s verified agent releases are: Salesforce Bridget (generally available in Regrello on 18 November 2026; expected in MuleSoft in December; no public price); Chargebee Arc (18 November, early access, no public price); Clay’s global agent (open beta, 19 November; 10,000 free beta credits on Enterprise, 1,000 on Growth, Launch and legacy self-serve); Hercules Agents (19 November; platform plans from $0 / $25 / $199 a month); Swan agents (17 November; Solo from $80–$100 a month); Gemini API Tool Retrieval (preview, 19 November; no separate list price); and Amazon Connect Resolutions (GA, 19 November; no Resolutions-specific price in the announcement).
Checked 20 November 2026 (Asia/Tehran) from official blogs, docs, pricing pages and Chargebee’s GlobeNewswire release. We did not benchmark these tools. Prices in USD before tax.
This is Sophrai’s weekly check of new AI agent tools, covering roughly 13–19 November 2026 — the days after our 13 November roundup. Every item below was re-read from an official primary source on 20 November 2026 (Asia/Tehran). We did not run or benchmark these products. Vendor figures for accuracy, latency, customer results or “100%” scores are labelled as vendor claims.
If you are comparing automation platforms rather than this week’s launches, we also published Make vs Activepieces (2026) today, alongside Zapier vs Activepieces and Zapier vs Make vs n8n. To sanity-check a stack, use the n8n cost calculator, the n8n workflow auditor and the AI SaaS pricing comparison.
Week at a glance (checked 20 Nov 2026)
| Tool | Type | Launched | Price / licence | Best for |
|---|---|---|---|---|
| Salesforce Bridget | Systems integration agent | GA in Regrello 18 Nov 2026; expected in MuleSoft in December | Not listed | Last-mile work inside browser-based enterprise apps |
| Chargebee Arc | Agentic finance platform | 18 Nov 2026, early access | Not listed / early access | Contract-to-revenue finance workflows with human approval |
| Clay global agent | GTM agent (open beta) | 19 Nov 2026 | Free beta credits: 10,000 (Enterprise) or 1,000 (Growth, Launch, legacy self-serve); then regular data credits | GTM teams already on a paid Clay plan |
| Hercules Agents | Self-improving business agents | 19 Nov 2026 | No separate Agents price. Platform: Free $0 (15 AI credits/mo), Pro $25/mo (50), Business $199/mo (50), Enterprise contact sales | Teams already building on Hercules |
| Swan agents | Named GTM agents | 17 Nov 2026 | Solo $100/mo ($80 yearly, 625 credits); Starter $200 ($160, 1,250); Growth $419 ($335, 2,000 credits + $99/$79 platform fee); Scale custom | RevOps and GTM teams that want a public self-serve price |
| Gemini Tool Retrieval | Gemini API preview | 19 Nov 2026 | No separate list price on the announcement | Developers with large tool libraries on Gemini 3.6+ or Gemini 4 |
| Amazon Connect Resolutions | Agentic operations (GA) | 19 Nov 2026 | Announcement points to Amazon Connect pricing; no Resolutions-specific list price in the post | Refunds, disputes, claims and similar high-volume cases |
1. Salesforce Bridget
Launched: Salesforce’s blog is dated 17 November 2026. It says Bridget is generally available in Regrello on 18 November and is expected to be generally available in MuleSoft in December. What it is: A headless systems-integration agent. Salesforce says you give Bridget sandbox access and a login; in about an hour (vendor claim) she explores a browser-based application, builds a library of governed, reusable MCP tools, and — after your team approves the work — locks the code so tasks run the same way every time. The post says she can diagnose interface changes and ‘self-heal’, runs on dedicated service accounts in your environment, and keeps a record of every action.
Price / licence (checked 20 Nov 2026): Salesforce does not list a price in the blog. Availability is described as general availability in Regrello on 18 November 2026, and expected in MuleSoft in December — ‘expected’ is the company’s word, not a date we can confirm yet.
What it means for n8n users: If you already orchestrate work in n8n and still have a last-mile step inside an app with no usable API, this is the class of tool that tries to close that gap. We found no official n8n connector for Bridget in this check. Treat any ‘self-heal’ or accuracy claim as unverified until you test it on your own applications.
The catch: The 100% accuracy figure, the 250 hours of benchmark testing, ‘zero model cost at run time’, and the Freeport-McMoRan anecdote (supplier onboarding ‘from weeks to days’) are Salesforce’s own claims in that post. We did not repeat the benchmark or speak to the customer. Price, contract terms and what ‘expected in MuleSoft in December’ will actually mean are not on the page.
Source: Salesforce blog: Meet Bridget
2. Chargebee Arc
Launched: Chargebee announced Arc on 18 November 2026 in a GlobeNewswire release sourced to Chargebee. The release says Arc is open for early access that day. What it is: An agentic finance platform that Chargebee says runs contract-to-invoice-to-revenue work as one traceable thread. At launch it ships three pre-built agents (contract ingestion, receivables, and revenue recognition) plus Agent Studio for custom agents. The release says every proposal is shown, cited and held for review before anything is committed, on a foundation of tasks, drafts, comments and citations.
Price / licence (checked 20 Nov 2026): No public list price appears in the press release. The product is described as early access. We also looked for a first-party Chargebee product or newsroom page for Arc on 20 November 2026 and did not find one (the paths we tried returned 404).
What it means for n8n users: If you already bill through Chargebee and want agents on contract intake, collections drafts or revenue-recognition proposals, this is the announcement to watch — but only after you get early-access terms in writing. We did not find an official n8n node for Arc. Chargebee’s older MCP is mentioned in the release as background, not as a price or an n8n integration.
The catch: We could not verify a Chargebee.com product page, pricing, SLAs, or how ‘early access’ is qualified. Quotes in the release are Chargebee’s. We did not use secondary news write-ups as sources.
Source: Chargebee’s GlobeNewswire release on Arc
3. Clay’s global agent (open beta)
Launched: Clay’s blog post is dated 19 November 2026. Official docs describe the global agent as in beta. What it is: An agent inside Clay that plans, builds and tests go-to-market plays across Search, Audiences, Workflows and Sequencer. Clay says it reads Knowledge Hub context first, tests on a sample of real records, shows an estimated cost on an approval card, and will not launch or send a campaign for you — it can draft, edit, spam-check and send one test email. Docs say Agent is in the left nav and also opens from the top bar, and that previous Sculptor entry points in Workflows, Search and the home screen now open the global agent (Sculptor stays in tables and Claygent Builder).
Price / licence (checked 20 Nov 2026): Docs: every eligible workspace gets free credits for the agent’s own work (reading the workspace, answering, planning and building) during beta — 10,000 on Enterprise and 1,000 on Growth, Launch and legacy self-serve plans. Free credits are spent first; enrichments, workflow runs and other downstream work always use regular data credits and actions. Trial workspaces get no free credits. Once regular credits run out, the agent stops even if free credits remain. Availability: beta on paid Enterprise, Growth and Launch plans, and for a limited time on legacy Starter, Explorer and Pro plans. Clay says free credits apply during beta and that pricing may change before general availability, to be posted on the same docs page first.
What it means for n8n users: If you already run Clay enrichments from n8n, the new agent lives inside Clay, not as an n8n node we could verify today. The useful part for an n8n-centred stack is the approval card and the fact that campaign send stays in your hands. Check the work: Clay’s own docs say results may not always be accurate while it is in beta.
The catch: We did not measure how fast free credits run down per message, or what a typical play costs in regular data credits after the bonus. ‘During the alpha, teams often used the global agent to learn Workflows faster’ is Clay’s wording, not our research.
Source: Clay blog (19 Nov 2026) and Clay docs: global agent
4. Hercules Agents
Launched: Hercules’ blog is dated 19 November 2026. What it is: Self-improving agents you describe as a job (‘build a chatbot for my customers’, ‘draft a quote when a request comes in’). Hercules says an agent uses your instructions, files and skills, connects through integrations or your own MCP server, and can live inside an app you built on Hercules. You can talk to it in the app or in Slack, Microsoft Teams, Google Chat, Linear, GitHub, Telegram, WhatsApp or X, or run it from a trigger, webhook or schedule. Auto-improve reads run history and proposes instruction changes, with credit caps you set; you can leave it suggesting, or switch it to apply fixes on its own. Every run is recorded with tools called and cost.
Price / licence (checked 20 Nov 2026): The launch post does not list a separate Agents price. Hercules’ pricing page the same day (monthly cards, checked 20 November 2026) lists Free $0 with 15 AI credits a month, Pro $25 a month billed monthly with 50 AI credits, Business $199 a month billed monthly with 50 AI credits, and Enterprise as contact sales. A yearly toggle is labelled ‘Save 15%’; we did not capture the yearly amounts. Top-ups are mentioned on Pro (‘Purchase AI credit top-ups’) without a public per-credit price on that page.
What it means for n8n users: If you already ship internal tools on Hercules, Agents are a way to put a guarded chat or triggered worker next to that app. For an n8n-centred stack, the official hook is MCP or a webhook trigger — we did not find an official n8n node. Watch Auto-improve credit caps if you turn on scheduled reviews.
The catch: We did not test Agents, did not confirm whether they run on the Free 15-credit plan, and did not capture yearly prices. The support-ticket example (an agent that reads docs, replies and closes tickets) is Hercules describing its own setup, not a third-party case study we verified.
Source: Hercules blog: Launching Hercules Agents and Hercules pricing
5. Swan agents
Launched: Swan’s changelog entry is dated 17 November 2026 (datePublished 2026-11-17). What it is: Named workers that own automations. Swan says each agent has a goal, tools, skills, memory and a scope; agents review results, improve what works, stop what does not, and can set reminders. Team and personal agents are supported. Existing automations were migrated into agents. The official how-agents-work docs sit alongside the changelog.
Price / licence (checked 20 Nov 2026): Swan’s pricing page on 20 November 2026 (not a separate Agents SKU) lists Solo $100 a month or $80 billed yearly (625 credits, 1 seat), Starter $200 or $160 yearly (1,250 credits, 1 seat), Growth $419 or $335 yearly (2,000 credits, 5 seats, plus a $99 / $79 yearly platform fee listed as $99 platform fee + $0.16 per credit, or $79 + $0.128 billed yearly), and Scale as custom. Extra seats are $20 a month. Unused monthly credits roll over up to 2× the plan; annual commitments can use the year’s credits any time in the term. Top-ups are listed at $0.37 (Solo), $0.30 (Starter) and $0.24 (Growth) per credit. Swan stops starting new chats and trigger runs when the balance drops below 10 credits, per the FAQ on that page.
What it means for n8n users: If you want a GTM agent product with a public self-serve price, Swan’s list is unusually complete compared with this week’s enterprise announcements. We did not find an official n8n integration page for Swan in this check. Confirm whether your work is seat-counted (people who chat or approve) before you model cost — alerts and views are described as free.
The catch: We did not run Swan or verify credit burn per task. The changelog does not restate prices; figures above are from the pricing page the same day. Whether ‘agents’ change credit use versus earlier Swan automations is not something we tested.
Source: Swan changelog: Meet Swan agents and Swan pricing
6. Gemini API Tool Retrieval
Launched: Google’s Gemini API blog is dated 19 November 2026. What it is: A preview feature in the Gemini Interactions API. Google says that when an agent has dozens or hundreds of tools, sending every full parameter schema on every request inflates input tokens and can hurt tool-selection accuracy. Tool Retrieval lets you add a tool_search tool and set defer_loading: true on tools you want held back: the model still sees each deferred tool’s name and description, and the API loads the full schema only when the model asks for that tool, inside the same request. The post says it works with Gemini 3.6 and later models, including all Gemini 4 models, in the Interactions API.
Price / licence (checked 20 Nov 2026): The announcement does not list a separate price for Tool Retrieval. We did not find a fee beyond ordinary Gemini API token billing on that page, and we are not inventing a token rate here — check Google’s current Gemini API pricing if you need official token prices.
What it means for n8n users: If you already call Gemini from n8n (HTTP Request or a community node) with a large tool list, this is aimed at you: fewer input tokens if the official behaviour matches the post. It is a preview. We have not wired it into an n8n workflow.
The catch: We did not measure token savings or tool-selection accuracy. Those benefits are Google’s stated design goals, not our test results. Preview availability, model list and API shape can change.
Source: Google blog: Tool Retrieval in the Gemini API
7. Amazon Connect Resolutions
Launched: AWS’s What’s New post is dated 19 November 2026 and announces general availability. What it is: An agentic operations solution in Amazon Connect for high-volume work such as refund requests, transaction disputes, fraud investigations and insurance claims. AWS says managers describe a process in natural language or upload standard operating procedures, and Resolutions generates a Work plan, asking clarifying questions about exceptions, decision rules and approval thresholds. Each incoming request becomes a Resolutions Case; AI agents gather information, apply policy, act across business applications and ask customers for missing details. Steps can run on prebuilt agents, agents configured in Amazon Connect Customer, or your own, with human specialists on approvals and exceptions. AWS says each outcome is verified and recorded.
Price / licence (checked 20 Nov 2026): The announcement does not list a Resolutions-specific price. It points readers to Amazon Connect documentation, a product webpage and a pricing page. We are not quoting a dollar figure we did not read from a full official rate card in this check.
What it means for n8n users: This sits in the contact-centre and back-office lane, not as an n8n replacement. If you already pass cases into Connect, it is worth reading the official docs before you build a parallel agent in n8n. We did not find an official n8n node for Resolutions in this check.
The catch: We did not verify runtime cost, required Connect editions, or regional availability. Time-to-resolution and cost-to-serve claims in the announcement are AWS’s. We did not test Resolutions.
Source: AWS What’s New: Amazon Connect Resolutions
Also this week
OpenAI introduced ChatGPT Wallet on 13 November 2026. OpenAI says it is a wallet that lets agents pay without access to the user’s actual card details and that requires a security check for every purchase, built with American Express, Chase, Mastercard and Visa. At launch it is described as rolling out in dots for ChatGPT Pro subscribers aged 18 and older in the US with eligible US-issued cards. We did not find a separate Wallet list price on that page, and we did not test it.
Superhuman announced on 19 November 2026 that it is expanding Go, including a personal AI assistant, a generally available Go platform SDK, and more than 20 partner agents to roll out over the coming weeks. Availability on that post is staggered: the personal assistant ‘will start rolling out over the next few months’; Go agents in Superhuman Docs are in open beta from 1 December for self-serve customers and in Q1 2027 for enterprise; new Mail features are in beta this month and generally available in early 2027. Superhuman’s ‘1 million monthly active users’ and ‘100 billion LLM calls a week’ figures are vendor claims. We did not treat these as generally available lead items.
Already covered on Sophrai
We are not re-leading items from the 13 November roundup (self-hosted Claude Managed Agents, n8n 3.2.0 pre-release, kagent 1.0, Fireflies Fred, Retell Tell-1, Meta Muse Ember, plus short notes on OpenAI Businesses in ChatGPT, Nectar and Microsoft Agent Framework python-1.25.0). Earlier weeks: 6 November and 30 October.
Separately this week we already published Trst AI agent for overseas sales and I connected Gemini’s free API to MausBot. Those are not repeated as lead items here. Browse more comparisons in the tools directory.
Which one should you try first?
It depends what you already pay for. If you are on a paid Clay plan, the global agent’s documented free beta credits (10,000 or 1,000) are the lowest-friction trial in this list — with the docs’ own warning to check its work. If you want a GTM agent with a public self-serve price, start with Swan’s published Solo or Starter figures and model credits honestly. If you build on Gemini and your tool list is getting large, Tool Retrieval is the one developer-facing preview worth reading this week. Bridget, Arc and Resolutions are enterprise announcements: ask for a quote and a sandbox, and do not budget from a price we could not find. Hercules Agents only make sense if you are already on that platform.
We did not rank these tools by quality. We did not run them.
What we couldn’t verify
- A public list price, contract pack or MuleSoft GA date for Bridget, and Salesforce’s own 100% accuracy / 250-hour benchmark.
- A first-party Chargebee Arc product page, early-access terms or price.
- Clay credit burn per typical play after the free beta grant, and any GA price (Clay says it may change).
- Hercules yearly card amounts, per-credit top-up rates, and whether Agents run on Free.
- Swan credit use per agent task, and whether agents are billed differently from earlier automations.
- Token savings from Gemini Tool Retrieval, and a separate feature price.
- A Resolutions-specific rate on Amazon Connect’s pricing page (the announcement only points to it).
- ChatGPT Wallet fees beyond a ChatGPT Pro subscription, and Superhuman Go GA dates that are still in the future.
- Official native n8n nodes for any of this week’s lead tools.
FAQ
What were the most important new AI agent tools this week?
The seven releases we verified from primary sources on 20 November 2026 are Salesforce Bridget (generally available in Regrello on 18 November, expected in MuleSoft in December), Chargebee Arc (announced 18 November, early access), Clay’s global agent (open beta, 19 November), Hercules Agents (19 November), Swan agents (17 November), Gemini API Tool Retrieval (preview, 19 November) and Amazon Connect Resolutions (general availability, 19 November). ChatGPT Wallet also began rolling out on 13 November.
How much does Clay’s global agent cost?
Clay’s docs, checked on 20 November 2026, say every eligible workspace gets free credits for the agent’s own work during beta: 10,000 on Enterprise and 1,000 on Growth, Launch and legacy self-serve plans. Those free credits are spent first; enrichments and workflow runs the agent sets up always use the workspace’s regular data credits and actions. Trial workspaces get no free credits. Clay says free credits apply during beta and that pricing may change before general availability.
How much does Swan cost?
Swan’s pricing page on 20 November 2026 lists Solo at $100 a month ($80 billed yearly) with 625 credits, Starter at $200 ($160 yearly) with 1,250 credits, Growth at $419 ($335 yearly) with 2,000 credits plus a $99 ($79 yearly) platform fee, and Scale as custom. Extra seats are $20 a month. Included credits are billed at $0.16 each on the monthly book ($0.128 billed yearly). Top-ups are listed at $0.37, $0.30 and $0.24 per credit on Solo, Starter and Growth. We did not run Swan, and the page prices the platform, not a separate add-on for agents.
Is Salesforce Bridget free, and did you test the 100% accuracy claim?
Salesforce’s 17 November 2026 blog does not list a price. It says Bridget is generally available in Regrello on 18 November and expected in MuleSoft in December. The 100% accuracy figure, the 250 hours of benchmark testing and the ‘zero model cost at run time’ line are Salesforce’s own claims from that post. We did not repeat the test.
What does Hercules Agents cost?
Hercules announced Agents on 19 November 2026 but did not publish a separate Agents price. The same day’s pricing page lists Free at $0 with 15 AI credits a month, Pro at $25 a month billed monthly with 50 AI credits, Business at $199 a month billed monthly with 50 AI credits, and Enterprise as contact sales. A yearly toggle advertised as ‘Save 15%’ was visible; we did not capture the yearly card amounts. We did not test whether Agents require a paid plan.
How did you check these tools?
We read each vendor’s own announcement, documentation, pricing page or press release on 20 November 2026 (Asia/Tehran). Benchmarks, accuracy figures and customer anecdotes are labelled as vendor claims; we did not run these tools or benchmark them. Cost examples marked as our arithmetic are calculated from the list prices shown, not quotes.
Sources and method
We fetched official pages on 20 November 2026 (Asia/Tehran) and saved raw dumps under our research notes. Primary sources: Salesforce Bridget blog; Chargebee Arc press release (GlobeNewswire, source: Chargebee); Clay global agent blog and Clay docs; Hercules Agents blog and Hercules pricing; Swan agents changelog and Swan pricing; Gemini Tool Retrieval blog; Amazon Connect Resolutions announcement; ChatGPT Wallet; Superhuman Go announcement. We did not use junk aggregators as sole sources. We did not benchmark these tools.
