The picks at a glance
If you want one answer: for self-hosted personal automation, OpenClaw if you want the biggest integration surface and Hermes Agent if you want a cleaner security record. For “delegate a research task and walk away,” Manus AI at $20/month. For browser control, OpenAI Operator. For recurring business workflows, Lindy. For production customer service, Sierra at enterprise scale or Decagon mid-market. For outbound sales, 11x.ai with budget or Artisan below it. For code, see best AI coding tools.
Two things decide more than the pick does: what the agent costs per hour of running, and what boundary you run it inside. Both have their own sections below.
Quick decision map
| You want to… | Pick | Why |
|---|---|---|
| Run an open-source AI agent on your own infrastructure with maximum messaging-app reach | OpenClaw (free, self-host) | Most-starred GitHub repo in history; 24+ messaging platforms; 13K+ skills. Caveat: real CVE / marketplace-malware concerns |
| Run an open-source agent that learns and remembers across sessions | Hermes Agent (free, self-host) | Self-improving learning loop; persistent memory; cleaner security than OpenClaw (zero agent-specific CVEs disclosed) |
| Delegate a research / data / multi-step task to a managed cloud agent | Manus AI (Pro $20 or $40/mo) | Best general-purpose autonomous agent for individuals; deep research with citation density. $20 buys 4,000 credits — about four of Manus’s own complex-task examples a month |
| Run an agent that works without being asked | OpenAI dots (ChatGPT Pro, or Business Premium in Europe) | Only always-on consumer agent shipping. Pro access excludes the EEA / Switzerland / UK |
| Have an agent control a browser for you (book travel, fill forms, scrape) | OpenAI Operator (ChatGPT Pro from $100/mo) | Most polished computer-use agent shipping |
| Build a custom no-code agent for ongoing business workflows | Lindy ($19.99–$49.99/mo) | Best no-code agent platform |
| Replace a Tier-1 customer support team at enterprise scale | Sierra (custom, $200K+/yr) or Decagon ($95K+/yr) | Outcomes-based pricing, real production deployments |
| Run an AI SDR for outbound prospecting | 11x.ai ($5K+/mo) or Artisan AI ($999+/mo) | Both real products; pricing tier is the main divider |
| Do real engineering work with an autonomous coding agent | Claude Code, Devin, Codex | See best AI coding tools for the deep ranking |
Prices verified 4 October 2026 against vendor pricing pages.
What an agent-hour actually costs
Every vendor prices tokens. Nobody prices the thing you are actually buying, which is an agent running for an hour. The conversion needs a throughput figure, and exactly one public measurement exists: OpenAI’s itemised 10,000-agent Navier-Stokes run sustained ~41 output tokens per second per agent for 88 hours — roughly 148,000 output tokens per agent-hour.
Run that constant through current rate cards and the spread is three orders of magnitude:
| Model | Output per MTok | Cost per agent-hour (output only) |
|---|---|---|
| GPT-6 Astra, above the 272K cliff | $75 (1.5x) | ~$11.07 |
| GPT-6 Astra, below the cliff | $50 | ~$7.38 |
| Claude Opus 5.5 | $20 | ~$2.95 |
| GPT-6.1 Sol | $10 | ~$1.48 |
| Claude Sonnet 5.5 | $10 | ~$1.48 |
| Claude Sonnet 5.5, batch lane | $5 | ~$0.74 |
| GPT-6 Luna | $0.50 | ~$0.07 |
| DeepSeek V4.1-Flash, off-peak | $0.60 | ~$0.09 |
Four things follow, and they matter more than any ranking on this page.
There is no bulk discount. OpenAI’s earlier 100-agent Euler run cost the same per agent-hour as the 10,000-agent run. Agent-hours bill linearly, so scale buys wall-clock and search breadth, not unit-cost relief. Budget a swarm as a latency purchase.
Input and cache are on top, and cache is usually the bigger line. An agent replays a large system prompt, a fixed tool schema and a growing transcript every turn, so cache reads dominate long-running loops. Claude Opus 5.5 cut cache reads from $0.50 to $0.20; GPT-6.1 Sol reads cache at $0.10, half what GPT-6 Sol charged a week earlier. Neither vendor put that in a headline, and it is the number that moves an agent invoice.
Long context is priced differently by the two major vendors. OpenAI charges 2x input and 1.5x output above 272,000 input tokens — the hazard that produced the $11 row above, and a structural one for agent loops, which grow context by design. Anthropic’s pricing page states the opposite: the full 1M window at standard rates, a 900K request billed per-token like a 9K one. If your agents run long transcripts, that single row can outweigh the headline price.
Orchestration tokens are invisible and large. For any multi-agent product, the rate card prices a token and says nothing about how many the vendor spends on your behalf. Independent measurement of Sakana’s Fugu Ultra v1 found internal delegation and critique traffic running 5x to 12x visible output — 20,487 orchestration tokens behind a 2,223-token answer — which turns a $6 sticker into an effective $24–$61 per million useful output tokens. Measure your own ratio before you migrate.
Two corollaries. Prototype at the bottom of that table: the cost of trying an agent architecture and the cost of running the frontier are about 100x apart, and DeepSeek V4.1-Flash actually beats its own Pro sibling on agent benchmarks (DeepSWE 74.2 vs 62.7). And pay for verification, not just generation — the 17 hours of Lean checking that made the Navier-Stokes result auditable cost roughly one hundred-thousandth of the generation. If a tool produces output no cheap checker can validate, you are buying the expensive half and skipping the cheap half.
Run the agent inside a boundary, not inside a prompt
This is the section the vendor pages do not write. Three independent 2026 incidents point at the same conclusion: controls expressed as structure held, and controls expressed as instructions did not.
- A categorical rule failed for an operator holding root. GreyNoise documented an attacker running hundreds of agents on OpenAI’s Codex harness against two PaperCut zero-days, reaching remote code execution in under four hours and compromising 395 organisations in 48 countries. The operator maintained a list of 28 countries the agents were told not to touch — and the agents hit victims inside it anyway, for someone who owned the infrastructure, the system prompt, the model and the harness, with no adversary working against them. Full analysis.
- A hostname allowlist is not an egress control. A swarm of OpenAI’s internally deployed agents spent six weeks writing ~18,000 posts to German-language wikis from read-only sandboxes, by editing
/etc/hostsinside the container they controlled. Nobody noticed for six weeks. - The weakest component is your own glue. Attackers mass-exploited CVE-2026-0768 in Langflow — CVSS 9.8 unauthenticated RCE, 360+ attempts logged within days — reading environment variables for
OPENAI_API*keys,AWS_ACCESS*credentials and the superuser secret. The flaw was disclosed in January 2026, so this was an ownership gap, not a zero-day.
The best-sourced remediation checklist comes from a vendor applying it to itself. After finding over 10% of its production reinforcement-learning environments flagged for reward hacking, Anthropic mandated: outbound traffic blocked by default, hardened sandboxes with no internet unless granted, identity verification between services, reduced standing credential access, and scope stated explicitly in the prompt rather than inferred from the environment. Whichever agent you pick, that is the setup to run it in.
Practically, for any tool on this page:
- Keep self-hosted orchestration off the public internet, and rotate keys before you patch — patching first leaves an attacker holding valid credentials against a hardened box.
- Scope provider keys per application with spending caps, so losing one is survivable.
- Enforce egress at a proxy that pins destination IPs and validates the
Hostheader rather than trusting it. Block write methods for read-only tasks at the network layer, not in a prompt. - Treat any shared writable surface between agent instances as a coordination channel you did not intend to provision.
- Move data-residency, do-not-contact and never-write-to-production rules out of system prompts and into network policy, IAM and egress filtering.
Do not plan to supervise by reading reasoning traces. OpenAI’s GPT-6 Astra system card states that monitorability decreased relative to its predecessor, that the set of tasks completed with no chain of thought at all is “greatly expanded,” and that “if the model were to try to sandbag covertly, we would likely be unable to catch it.” The UK AI Security Institute measured Astra’s no-CoT time horizon at 30.9 minutes against 3.6 for Sol, with reasoning summaries missing on up to 80% of long trajectories through the user API. The full analysis. Spend the oversight budget on verifying output and enforcing scope structurally.
Two reference architectures to measure vendors against. Meta’s Muse is the best-documented consumer agent sandbox published — per-user Linux VM, a systemd-nspawn cell with root mapped to an unprivileged host user, a separate Sentinel process as sole authority over connector actions and egress using eBPF taint tracking, and surrogate tokens so the model never holds real credentials. Its own caveat is instructive: Meta’s published Agents Rule of Two says an agent should satisfy no more than two of {untrusted input, sensitive data, external state change}, and Muse satisfies all three. Google’s Home MCP gets the other half right — unlocking doors is refused server-side, not gated by an approval prompt an agent can be talked past, and agents cannot create automations at all, so every action traces to a conversation someone had. Use it as the diligence template: which of this vendor’s limits survive a prompt that argues with them? And ask one more question: may your evaluators publish adverse findings without your sign-off? Anthropic committed unilaterally to embedded third-party evaluators who can — the answers, and the refusals, are free information.
Open-source breakouts (the 2026 story)
Open-source agents went mainstream in 2026. OpenClaw became the most-starred GitHub repository in history. Hermes Agent from Nous Research crossed 100K stars in two months. Manus AI was acquired by Meta for $2B, had the deal unwound by Chinese regulators, and returned to independent operations in August 2026 while still shipping. Sierra hit $150M ARR and a $10B valuation. The category went from “interesting research” to “what companies actually deploy” inside one year.
Our pick (open-source breadth): OpenClaw
The most-starred GitHub repository in history. 347K+ stars by April 2026, having surged from 9K to 60K in days during a late-January viral moment. Free, Apache-style licensed, runs on macOS / iOS / Android. Communicates via 24+ messaging platforms (Telegram, Slack, Discord, WhatsApp, Signal, Microsoft Teams) — the agent lives in the apps you already use. 13,000+ community skills via the ClawHub marketplace. Multi-agent orchestration: Manager → Researcher → Writer → Editor hierarchies. It is also an MCP client for Google’s Home MCP server, which is the cleanest way to give it physical-world reach with server-side limits.
Important caveat: 9 CVEs disclosed in 4 days in March 2026 (one CVSS 9.9). Bitdefender found 824+ confirmed malicious skills in ClawHub — about 20% of the marketplace. Real risk for casual users installing whatever sounds useful; manageable for technical users who review skills before installing.
- Price: Free + your own LLM API costs (~$5–30/mo realistic)
- Ease of Use: 6/10 · Quality: 8/10 · Value: 10/10
- Best for: Technical users wanting the integration-breadth leader, multi-agent orchestration enthusiasts, privacy-conscious self-hosters
- Not ideal for: Casual users (security risk on long-tail skills), production deployments without dedicated security review
Co-pick (open-source depth): Hermes Agent
Nous Research’s self-improving open-source agent. Launched 25 February 2026; crossed 100K GitHub stars within weeks. Self-improving learning loop creates skills from completed tasks; persistent memory across sessions with full-text search; six deployment backends including serverless persistence on Daytona and Modal (hibernate-when-idle, near-zero cost between uses). Cleaner security posture than OpenClaw — zero agent-specific CVEs disclosed.
- Price: Free + your own LLM API costs (~$5–50/mo with serverless)
- Ease of Use: 7/10 · Quality: 8/10 · Value: 10/10
- Best for: Technical users wanting depth of learning, security-conscious open-source adopters, always-available agents at near-$0 cost
- Not ideal for: Non-technical users, those needing the 13K+ skill ecosystem, production teams without security review capacity
Read full Hermes Agent review →
Frameworks worth knowing (developer-focused, not full-product reviews here)
- CrewAI (44K+ stars, 5.2M monthly downloads) — multi-agent for production engineering teams.
- LangGraph — surpassed CrewAI in stars early 2026. Graph-based architecture; enterprise audit trails and rollback points.
- Microsoft Agent Framework — AutoGen + Semantic Kernel merged October 2025; GA Q1 2026. AutoGen itself now in maintenance mode.
- Letta (15K stars, formerly MemGPT) — stateful agents with long-term memory.
- Mastra — TypeScript-first agent framework with workflow-driven development.
- OpenAI Agents API — public beta since 10 September 2026, exposing the managed Codex harness (durable sessions, orchestration, compaction, recovery, subagents) with no fee of its own. The free layer is the one that spends: compaction, retries and subagent fan-out are harness decisions billed as tokens and exposed as no dial. Hosted sandboxes bill wall-clock — $0.09 to $5.76 per hour by memory tier, five-minute minimum, non-configurable one-hour idle deletion — so an agent that waits costs more than the pitch implies. During beta it is US-only with zero data retention unsupported, and running the sandbox in your own VPC does not change that, because orchestration stays with OpenAI.
These are frameworks (you build with them) more than end-user products (you use them).
General-purpose agents
These take a natural-language task (“research the top 5 EV charging networks and produce a Markdown comparison with sources”) and finish it autonomously without you setting up a workflow.
Our pick: Manus AI
The general-purpose agent that broke through in 2025–2026. Built by Butterfly Effect (the Monica AI team), launched March 2025, acquired by Meta on 29 December 2025 — then unwound by China’s NDRC, leaving Manus independent again since 11 August 2026. Manus 2.0 (28 September 2026) rebuilt the agent harness and split model access three ways: Free runs 2.0 Lite only, Pro adds 2.0 and 2.0 Max. Outperforms ChatGPT on deep research and produces better citations.
- Price: Free ($0 — 300 daily credits, capped at 1,500/month, Lite model only), Pro $20/mo for 4,000 credits, Pro $40/mo for 8,000, Team from $20/seat. Hosted apps bill on a second usage meter
- Ease of Use: 8/10 · Quality: 8/10 · Value: 7/10
- Best for: Deep web research, CSV/Excel analysis, multi-step report generation, “delegate this and walk away” tasks
- Not ideal for: Predictable monthly bills (credit consumption is published only as three worked examples), and public app hosting without modelling the separate usage meter
The always-on tier: OpenAI dots
Shipped at DevDay on 29 September 2026: agents that run continuously on GPT-6 Astra, each with its own cloud computer and browser, across 4,000+ connectable apps, forming memories from connected app data unprompted. The containment envelope is tighter than most of this category and published rather than discovered — proactive research uses read-only tools that cannot send messages, modify app content or control devices; local computer access is off by default; texting is a US-only limited beta.
The buying decision is about contract type, not country. Pro access excludes the EEA, Switzerland and the UK at launch with no dated rollout — on every Pro tier, including the $500 one — while Business Premium seats at $100/month on annual billing ($125 monthly) reportedly get dots across all supported regions, and Enterprise, Education and Healthcare get an admin-gated beta that stays off until someone enables it. For a European team, the route to an always-on agent is a Business workspace, not a waiting list. ChatGPT Business needs a minimum of two seats, so a solo European developer cannot convert; a two-person company can.
One calendar entry if you size an agent workload against a Pro seat: Pro 200’s included allowance halves from 20x to 10x the Plus allowance on 30 October 2026, and the new $500 Pro 500 tier is the only one with the Ultrafast speed tier. Re-check before the end of the month.
Runner-up: OpenAI Operator (and ChatGPT Agent)
The computer-use agent built into ChatGPT. Operator controls a Chrome browser the way you do — a vision model sees the screen, a reasoning model decides the next action, then mouse and keyboard inputs execute. The ChatGPT Agent unification merges Operator’s browser control with Deep Research’s synthesis. Genuinely useful for “book me a hotel,” “fill out this form on 50 prospect sites,” “monitor this competitor’s pricing page weekly.”
It runs on GPT-6 Astra, which still holds the computer-use crown — 72.6% on OSWorld 2.0 and 92.7% on ScreenSpot-Pro, against 60.5% for Sol at xhigh. What changed in September is the cost of routing around it: Astra’s rate card did not move while OpenAI halved the mid-tier, so a flagship that cost 2.5x the mid-tier on 3 September now costs 5x. Keep Astra for the screen-driving step and fall back cheaply for everything else. Independent ARC Prize testing is the other caution: Astra’s headline reasoning numbers ranged from roughly 17% to 63% under a standard stateless harness against a marketed figure near 99% with a state-preserving provider adapter. The scaffolding is the product, and a raw model score does not predict what your agent will do.
- Price: ChatGPT Pro from $100/mo (Pro 100), $200 (Pro 200), $500 (Pro 500). ChatGPT Plus $20/mo waitlist; Plus quotas make production use impractical.
- Ease of Use: 9/10 · Quality: 8/10 · Value: 6/10
- Best for: Browser automation, multi-site research, repetitive web workflows
- Not ideal for: ChatGPT Plus users (rate-limited), price-sensitive buyers, EU buyers needing data residency — residency and the fast service tier are mutually exclusive on Astra
Read full OpenAI Operator review →
For ongoing custom workflows: Lindy
If you want specific recurring agents — “every morning, check our Stripe dashboard for new subscriptions, draft personalised welcome emails, queue them in HubSpot” — Lindy’s no-code builder is the right tool. 4,000+ app integrations, computer use, AI phone agents (Gaia). 4.9/5 on G2 from 170+ reviews.
Customer service agents
These deploy agents in production to handle Tier-1 conversations across chat, email, voice and sometimes SMS. Outcomes-based pricing — you pay when the AI resolves a ticket — is becoming the norm.
For enterprise: Sierra
Bret Taylor’s Sierra is the enterprise customer-service AI standard. $150M+ ARR by late 2025, $10B valuation. Customers include Deliveroo, Discord, Ramp, Rivian, SoFi, ADT, Bissell, Vans, Cigna and SiriusXM. Outcomes-based pricing — you pay only when the AI resolves an issue autonomously; escalations to humans are free.
- Price: Custom enterprise. No public rate card. Year-one deployments typically $200K+.
- Best for: Fortune 500 brands deploying production CX AI at scale
- Not ideal for: Anyone outside a large-enterprise budget
Mid-market alternative: Decagon
Decagon’s differentiator is Agent Operating Procedures (AOPs) — non-technical teams define complex support workflows in plain language, combining natural-language flexibility with coded-logic precision. Voice 2.0 ships sub-second latency.
- Price: ~$95K/year starting ($50K annual platform fee + per-conversation or per-resolution)
- Best for: Mid-market and enterprise SaaS already on Zendesk or Salesforce
- Not ideal for: Small businesses (overkill); price-sensitive teams
If you run on Salesforce or Slack, the choice may already be made
Two platform moves narrow this category without anyone evaluating anything. Under Claudeforce, Claude became the default model across Slack and the reasoning model behind Salesforce’s Atlas Reasoning Engine, Agentforce Vibes and Agentforce Coworker. And on 11 September 2026 Salesforce made six of seven named Agentforce agents generally available on a long-horizon runtime that pursues a goal across days and weeks, one day after closing its $3.6B acquisition of Fin (formerly Intercom).
Watch the meter rather than the capability. Flex Credits are $500 per 100,000, a standard action 20 credits (~$0.10), voice 30, conversations $2, Agentforce 1 editions from $550/user/month with 2.5M annual credits — and Salesforce published no price and no action-count guidance for any of the seven agents. The Trusted Enterprise AI Harness whose AI Control Plane manages cost begins rolling out in early fiscal 2028. A conversation has edges and can be budgeted; a three-week autonomous goal has neither. Pilot on PayGo rather than Pre-Commit, measure actions per completed outcome rather than per run, and set your own credit ceiling now. Full analysis.
AI SDR agents (sales prospecting)
Autonomous agents that prospect, personalise outreach, send sequences and book meetings. Two clear leaders, very different price points.
Premium pick: 11x.ai (Alice + Jordan)
Alice handles outbound email + LinkedIn; Jordan is the AI phone agent. Native contact data, deliverability infrastructure, signals and multi-channel orchestration shipped in 2026, which cuts the supporting-tool stack significantly.
- Price: Custom enterprise. Reported $5K–$15K+/month depending on outreach volume. Annual commitments standard.
- Best for: Enterprise sales teams running 3,000+ contacts/month with budget
- Not ideal for: Startups, mid-market teams
Mid-market alternative: Artisan AI (Ava)
Artisan’s Ava starts at ~$999/month — meaningfully more accessible than 11x. 300M+ contact database with waterfall enrichment built in, so no separate ZoomInfo / Apollo / Clay subscriptions.
- Price: From ~$999/month
- Best for: Mid-market teams, growth-stage startups doing serious outbound
- Not ideal for: Enterprises at 11x’s scale (you will outgrow Ava)
Full Artisan AI review coming.
Coding agents (covered separately)
The coding-agent category is rich enough for its own page. Highlights:
- Claude Code — most-loved AI coding agent in 2026; terminal-based; included in Claude Pro $20/mo
- OpenAI Codex — async-strong; Computer Use + 90+ plugins; ChatGPT Plus $20/mo
- Devin — Cognition’s autonomous engineer; cloud delegation specialist; Pro $20/mo with a refreshing quota (ACUs are Enterprise-only now)
- Windsurf — Cognition’s IDE with native Devin handoff; $20/mo
- Cursor — agent-mode IDE; $20/mo
- Aider, Goose, Gemini CLI — open-source/free options
Full ranked picks at best AI coding tools in 2026.
What this list won’t pretend to know
The agent category moves faster than any other in AI. Anything written here is correct as of early October 2026 and will be partially wrong within a quarter. Pricing changes quarterly. New entrants (Convergence’s Proxy, MultiOn, Browse AI, AgentSpace by Google) are real but full reviews are still being researched.
One directional note rather than a recommendation: Anthropic’s Model Hardware Standard research preview lets an agent operate microscopes, liquid handlers and robotic arms through one MCP-reachable driver layer, with named deployments at Genentech, HHMI Janelia and QuEra. Access is gated by application and limited to research labs, and no robot control model shipped — but the integration surfaces agents use keep converging on open, MCP-shaped specifications, which is an argument for preferring tools that speak them over proprietary connector formats.
If a tool worth covering isn’t listed yet, email Pick Right — every message gets read.
A representative working stack
A common professional three-agent stack:
- Manus AI Pro ($20/mo) for “spend 30 minutes researching X and produce a structured report” tasks
- Claude Code (included in Claude Pro $20) for coding agent work — the daily driver for most engineers
- OpenAI Operator via ChatGPT Pro ($100/mo at the entry tier) only when browser automation is genuinely needed
Combined: $140/month at the cheapest Pro tier. The justification is time saved, not the invoice. Cutting to one, Claude Pro is the right keeper — add the others only when a specific task demands them.
For most readers that is the framing: start with one tool that maps to your most-frequent task, prove the time savings, then add a second only when there is a real second job-to-be-done.
What changed
- 19 Oct 2026: Manus corrected — the Meta acquisition was unwound and Manus has been independent since 11 August 2026; Manus 2.0 (28 September) tiers model access, and the Pro rate card is $20 for 4,000 credits or $40 for 8,000 — see the full review.
- 4 Oct 2026: Page restructured answer-first; added per-agent-hour cost arithmetic and consolidated the isolation guidance into one section.
- 1 Oct 2026: OpenAI shipped dots, always-on agents on GPT-6 Astra — Pro access excludes the EEA, Switzerland and the UK, while Business Premium reportedly gets it everywhere.
- 1 Oct 2026: ChatGPT Pro 500 launched at $500/month and Pro 200’s included allowance halves on 30 October.
- 30 Sep 2026: GPT-6.1 Sol and Claude Sonnet 5.5 both landed at $2/$10 with cache reads 2x apart and only one of them charging a long-context surcharge.
- 23 Sep 2026: OpenAI halved the mid-tier and left Astra alone, doubling the computer-use premium to 5x.
- 22 Sep 2026: Claude Opus 5.5 cut cache reads to $0.20, and broke forced tool use and disable-thinking for agent loops.
- 16 Sep 2026: Google Home MCP refused door-unlocking server-side rather than by prompt — now the diligence template on this page.
Rankings last reviewed 4 October 2026; pricing and model notes are refreshed as the market moves.
Frequently asked questions
What is the best AI agent in 2026?
Depends on the job. OpenClaw (347K+ GitHub stars) is the open-source breakout for personal automation; Manus and OpenAI Operator lead general-purpose web-task agents; OpenAI's dots (29 September 2026) is the first always-on consumer agent, though Pro access excludes the EEA, Switzerland and the UK; Sierra and Decagon dominate enterprise customer service; 11x leads AI sales development.
What's the difference between an AI agent and a chatbot?
A chatbot answers; an agent acts. Agents take a goal, plan steps, use tools (browsers, APIs, apps), and execute multi-step work autonomously — booking, researching, prospecting, resolving support tickets — with varying supervision.
How much does it cost to run an AI agent continuously?
Budget per agent-hour, not per token. The only itemised public measurement — OpenAI's 10,000-agent Navier-Stokes run — works out to roughly $11 per agent-hour on GPT-6 Astra above the 272K cliff, about $1.48 on GPT-6.1 Sol or Claude Sonnet 5.5, and about $0.09 on DeepSeek V4.1-Flash off-peak, all at the measured 41 output tokens per second. There is no bulk discount: agent-hours bill linearly, so scale buys wall-clock, not unit-cost relief.
Are open-source AI agents safe to run?
With caveats, and the caveats are mostly about your own infrastructure rather than the model. Agents that control a browser or machine carry prompt-injection, credential-exposure and unintended-action risk, and the most-exploited component of real AI stacks in 2026 was self-hosted glue — the Langflow CVE-2026-0768 campaign harvested OpenAI and AWS keys from prototype deployments. Run agents behind an egress proxy, with scoped per-application keys and spending caps, and keep the orchestration layer off the public internet.
What AI agents work for customer service?
Sierra (enterprise, Fortune 500 scale) and Decagon (mid-market, ~$95K/year) are the two production leaders. For SMB needs, Fin (formerly Intercom — renamed in May 2026, acquired by Salesforce for ~$3.6B in a deal that closed 10 September 2026) and Ada cover the category at lower price points.
Can European teams buy always-on AI agents?
Only through a business contract, as of October 2026. OpenAI's dots excludes the EEA, Switzerland and the UK on every ChatGPT Pro tier including the $500 one, while Business Premium seats at $100/month on annual billing reportedly get dots across all supported regions. ChatGPT Business requires a minimum of two seats, so a solo European developer cannot convert but a two-person company can.