Vertical AI / Agent Architecture
The Agent-First 10: Vertical AI Agents Deployed in the Real World
Most products marketed as agents in 2026 are still copilots wearing a new label. Here is the architecture test that separates the real ones, and ten vertical agents that pass it in production.
An agent is not a smarter chatbot. A chatbot answers a question and the loop ends. A copilot rides alongside a person, suggesting while a human stays on every keystroke. A truly agentic system runs its own loop: it plans, calls tools, holds context across steps, makes judgment calls, and hands back finished work. That last category is small, even though the label is now stamped on almost everything. The fastest way to cut through the marketing is to stop watching demos and start reading architecture. What follows is a field guide to ten vertical agents already running in production, and the test that tells you which ones are real.
PILLAR 01The line between a copilot and an agent
Think of three rungs on a ladder. The chatbot lives in a single turn: you ask, it answers, the loop ends. The copilot rides alongside a human, drafting and suggesting while a person stays on every keystroke. The agent runs its own loop: it plans, calls tools, holds context across steps, exercises judgment, and produces a finished work product that is judged on the output, not on how pleasant the conversation was.
That ladder gives you a six-point test, and it is worth slowing down on, because it is the same test we apply to every agent we evaluate. First, an autonomy loop instead of a single turn: the system decides its own next step rather than waiting to be told. Second, real tool use: it writes to the CRM, opens the pull request, books the slot, files the ticket, instead of describing what you should go do. Third, persistent memory: it carries context through a multi-step task rather than forgetting between prompts. Fourth, commit-point design: irreversible actions pause for a human by deliberate construction, not by accident. Fifth, outcome ownership: it is judged on the work product, not on how fluent the conversation felt. Sixth, agent-first architecture: it was built native to this era rather than retrofitted onto software designed for human seats.
Hold any product against those six and the copilots fall away within minutes. Most tools pass two or three. The ten below pass all six in at least one core workflow, which is a far higher bar to clear than a polished demo. A demo flatters the model. Architecture exposes it.
PILLAR 02Why agent-first is the real wedge
The model is the commodity layer. The defensible value sits underneath it in three tiers: proprietary data, deep integrations into the systems where work actually happens, and codified standard operating procedures that encode how a vertical really runs. Those tiers take tens of thousands of hours to build, which is exactly why incumbents struggle to retrofit them onto seat-based software.
This is why the phrase agents are eating SaaS keeps surfacing, and why we think it is only half right. The seat-based software model sold you access to a tool and left the work to you. The agent model sells the work itself, which is why pricing across the category is shifting from per-seat to per-outcome. That shift is not theoretical. As of the second quarter of 2026, roughly half of enterprises already run AI agents in production and another quarter are actively scaling them. Gartner expects task-specific agents to appear in around forty percent of enterprise applications by the end of the year, up from under five percent a year earlier.
The uncomfortable part for incumbents is that the three tiers do not transfer. A company can bolt a chat box onto its product in a weekend. It cannot manufacture a decade of proprietary workflow data, deep integrations, and codified procedure on the same timeline. That is the gap the agent-first challengers are running through, and it is why the winners in each vertical below are mostly new names rather than the established suite vendors. The model is rented from a handful of labs. The moat is built by hand, in one vertical at a time.
PILLAR 03The ten vertical agents in production
Read these not as a ranking but as a portfolio: the same architectural bet placed in ten different verticals. Each entry follows the same four beats, the job to be done, the deployed agent, the action that makes it genuinely agentic, and where the human commit point sits.
The ten split into two tiers, and the split is instructive. The first tier is mature: software engineering, customer service, legal, and clinical documentation have crossed from pilot into daily dependence, with revenue and usage figures that are no longer rounding errors. The second tier is still proving itself: sales development, security triage, and parts of finance work well at the top of the funnel but have not yet earned the right to act unsupervised on the highest-stakes decisions. Reading the list this way matters for buyers, because the maturity of the vertical, not the polish of the vendor, tells you how wide a commit point you can safely leave open today.
These agents take a written ticket and return working, tested code, operating like a junior engineer who never sleeps. They read the repository, plan the change, write and run code in a sandbox, fix their own failing tests, and open a pull request for a human to review.
- Agentic, not a chatbot: A chatbot suggests code for you to paste. This agent runs the loop itself, planning the change, editing files, executing tests, and revising until they pass, then opening the PR. It acts in the codebase instead of describing what to type.
- What it does: Turns a plain-language task into a tested, reviewable code change across a real codebase.
- How it works: Ingests the repo for context, breaks the work into a plan, executes in an isolated environment with terminal and browser access, and iterates against test results until they pass.
- Value created: Compresses routine work, bug fixes, migrations, boilerplate, from days to minutes, freeing senior engineers for architecture instead of toil.
- Commit point: The pull request. A human reviews and merges, so nothing reaches production unseen.
- Deployed: Devin from Cognition, billed as the first AI software engineer, with Claude Code, Cursor, and Replit Agent.
These agents resolve customer issues end to end instead of routing them. They hold a real conversation and then take the action the customer actually wanted, like issuing a refund or changing an order, rather than handing off a transcript.
- Agentic, not a chatbot: A chatbot answers FAQs and hands you off. This agent reasons over the customer's real account and executes the resolving action, the refund or the change, through live system calls, then confirms it worked.
- What it does: Handles inbound support across chat, email, and voice and closes the ticket rather than deflecting it.
- How it works: Connects to the knowledge base and back-end systems for orders, billing, and CRM, reasons over policy, and executes the resolving action through those integrations.
- Value created: Resolves a large share of tickets with no human, cuts response time to seconds, and holds quality steady at volumes a human team could never staff.
- Commit point: Escalation to a human on anything outside policy or below a confidence threshold.
- Deployed: Sierra, which reached one hundred million dollars in revenue in seven quarters, plus Decagon, Ada, and Intercom's Fin.
These agents run the repetitive top of the sales funnel: finding the right accounts, researching them, and sending personalized outreach at a scale no human team can match. The honest read is that they own prospecting, not closing.
- Agentic, not a chatbot: A chatbot writes a message when asked. This agent decides who to target, runs multi-touch sequences over days, watches for replies, and adapts its next move on its own, with no prompt per step.
- What it does: Builds target lists, researches prospects, writes and sends multi-step outreach, and books qualified meetings.
- How it works: Pulls from data providers and intent signals, personalizes messaging per prospect, and runs sequences across email and social, syncing everything back to the CRM.
- Value created: Gives a small team the reach of a large one and keeps pipeline flowing around the clock, freeing reps for the conversations that actually convert.
- Commit point: The human rep takes over at the qualified-meeting stage, and outbound is reviewed before live send in most deployments.
- Deployed: 11x with Alice, across hundreds of companies, and Artisan with Ava.
Legal agents take on the document-heavy core of legal work, reading large volumes of contracts and case material and producing first-pass drafts and analyses that a lawyer reviews. They are built around how firms actually operate, not as a generic assistant.
- Agentic, not a chatbot: A chatbot answers one legal question. This agent works an entire matter, opening and cross-referencing dozens of documents and holding that state as it assembles a cited work product.
- What it does: Performs due diligence, contract review, legal research, and drafting across a matter.
- How it works: Operates over firm and matter documents with retrieval and reasoning tuned for legal work, returning citations and structured output a lawyer can verify.
- Value created: Collapses the hours associate teams spend on review and research while raising consistency across a matter.
- Commit point: The supervising lawyer reviews and signs off; the agent never files or advises on its own.
- Deployed: Harvey, which also publishes a Legal Agent Benchmark to measure agents against real lawyer tasks.
Clinical agents split into two jobs: listening to a doctor-patient visit and writing the clinical note, and conducting lower-risk patient outreach by phone at scale. Both operate in a domain where the commit point is non-negotiable.
- Agentic, not a chatbot: A chatbot waits for you to type. This agent listens to a live visit and structures it into the record, and on calls it runs a goal-directed conversation that branches on what the patient says.
- What it does: Generates structured clinical documentation from ambient conversation, and runs follow-up and intake calls with patients.
- How it works: Transcribes and structures a visit into the record format clinicians require; for outreach, it uses safety-constrained voice models built specifically for healthcare.
- Value created: Returns hours of documentation time to clinicians each day, reduces burnout, and extends a care team's reach without adding staff.
- Commit point: The clinician edits and signs every note; clinical-judgment calls stay with licensed staff.
- Deployed: Abridge, valued above five billion dollars, and Hippocratic AI, which has run more than one hundred million patient interactions.
Field-service agents answer the phone for home-services businesses, plumbing, HVAC, junk removal, that lose revenue every time a call goes to voicemail. The agent picks up, qualifies the job, and books it straight into the schedule.
- Agentic, not a chatbot: A chatbot needs a keyboard. This agent answers a live phone call, holds a spoken back-and-forth in real time, checks the calendar, and commits the booking, with no human in the seat.
- What it does: Handles inbound and outbound calls, answers questions, and books and dispatches jobs.
- How it works: A voice agent wired into the company's scheduling and CRM systems, trained on its services, pricing, and booking rules.
- Value created: Captures revenue that used to leak through missed calls and after-hours gaps, with no added headcount.
- Commit point: The booked job and dispatch are visible to the operator, who controls scheduling rules and overrides.
- Deployed: Avoca, on track to book around one billion dollars in jobs in 2026 across more than eight hundred operators including national home-services brands.
Finance agents take on the multi-step, rules-heavy back-office work that consumes finance and operations teams: matching invoices, reconciling accounts, and processing compliance cases. Tight permission scoping is what keeps them safe.
- Agentic, not a chatbot: A chatbot explains the policy. This agent executes the workflow, pulling and matching invoices, flagging exceptions, and writing entries across systems while leaving an audit trail behind it.
- What it does: Runs accounts payable, reconciliation, expense review, and KYC and AML case work end to end.
- How it works: Connects to ERP, banking, and ledger systems, applies the organization's policies as codified rules, and produces an auditable trail of every action.
- Value created: Removes hours of manual data work, catches errors and anomalies earlier, and keeps a clean audit record, which is the real buying gate in regulated finance.
- Commit point: Payments, write-offs, and compliance decisions above a threshold route to a human approver.
- Deployed: Ramp's finance agents and a growing set of banking-grade workflow agents built for audit trails.
Security agents take the flood of alerts that overwhelms a security operations center and do the first pass, gathering context, correlating signals, and triaging what matters so human analysts spend their time on real threats rather than noise.
- Agentic, not a chatbot: A chatbot tells you what an alert means. This agent investigates it, pulling logs, pivoting on what it finds, and correlating signals into a triaged verdict, choosing its own path through the data.
- What it does: Investigates and triages security alerts, assembling the context an analyst would otherwise gather by hand.
- How it works: Pulls from logs, threat intelligence, and security tooling, adapts its investigation path to what it finds, and returns a prioritized, explained assessment.
- Value created: Cuts the time to investigate an alert dramatically and lets lean teams cover far more ground, easing analyst burnout and alert fatigue.
- Commit point: Containment and remediation stay with a human analyst; the agent recommends, the human acts.
- Deployed: the SOC triage agent pattern now moving into production across enterprise security teams.
Recruiting agents run the funnel from sourcing through scheduling as one continuous motion: finding candidates, screening them, and booking the interview, while holding borderline cases for a human recruiter to judge.
- Agentic, not a chatbot: A chatbot answers candidate questions. This agent runs the funnel, searching databases, screening and ranking, and booking interviews, deciding which candidates advance and which to hold for a human.
- What it does: Sources candidates, screens and ranks them against the role, and schedules interviews.
- How it works: Searches large candidate databases, integrates with applicant tracking systems, runs structured screening over chat or voice, and triggers scheduling automatically for strong matches.
- Value created: Turns weeks of manual sourcing and back-and-forth into hours and gives small talent teams enterprise reach. More than half of talent leaders plan to deploy autonomous agents in 2026.
- Commit point: Borderline candidates are held for recruiter review, and the hiring decision always stays human.
- Deployed: Pin, which runs a source-to-schedule loop across more than one hundred ATS platforms, and Perfect, which auto-triages and escalates by design.
Coordination agents do the deceptively hard work of getting the right people into the right meeting at the right time, negotiating across calendars, constraints, and time zones. This is the layer every other agent on this list eventually has to call.
- Agentic, not a chatbot: A chatbot can draft an email proposing times. This agent negotiates across many calendars, holds tentative slots, resolves conflicts, and commits the final event, managing a stateful, multi-party process.
- What it does: Negotiates, holds, and commits time across people and systems, and orchestrates the multi-step workflows that end in a booking.
- How it works: Reads availability and constraints across calendars, proposes and negotiates options, and writes the confirmed event back, with a human gate on anything sensitive.
- Value created: Removes the hidden tax of scheduling friction from every team, and grows more valuable as more agents need to book time, a position that compounds across the agent economy.
- Commit point: The reversible line is built in. Proposing times runs automatically; confirming or declining a meeting pauses for the human when it matters.
- Deployed: Lindy for broad operations automation, and ANCI, whose agent Zara coordinates scheduling natively rather than as a calendar add-on.
PILLAR 04The pattern across all ten
Read the list as a system and three things repeat. Every winner is agent-first, not AI poured over old software. Every winner's moat is codified procedure, not raw model access. And every winner that earns trust is designed around a clear commit point, the moment an irreversible action pauses for a human before it executes.
The question is no longer whether an agent can act. It is whether you can see, and control, the moment it commits.
Commit points are where most agent programs will live or die, so it pays to be precise about them. We sort every action an agent can take into two buckets. Reversible actions, like drafting a reply, proposing a shortlist, or suggesting a meeting time, can run fully autonomously, because a mistake costs a click to undo. Irreversible actions, like sending the email, declining the candidate, moving the money, or merging to production, are the ones that need a gate. The design skill is not adding human review everywhere, which kills the very speed that made the agent worth deploying. It is putting the review in exactly one place: the threshold between reversible and irreversible.
For leaders, that taxonomy turns into a buying discipline. Stop evaluating agents by how they perform in a scripted demo, where every input is friendly and every edge case is hidden. Evaluate them by their work product over a week of real tasks, and by exactly where the commit point sits. An agent with no commit point is a liability dressed as productivity. An agent with a well-placed one is the first piece of software you can genuinely treat like a hire.
It is worth being honest about why so many agents stall short of this standard. The distance from a convincing demo to a production system that holds up at ninety-nine percent accuracy is not incremental, it is exponential. Demos run on happy paths. Production runs on edge cases, malformed inputs, ambiguous instructions, and the long tail of exceptions no script anticipates. This last mile is exactly where codified procedure earns its cost, because the SOPs encode how a vertical handles the cases that break a general-purpose model. The agents that win are not the ones that demo best. They are the ones that fail gracefully, escalate cleanly, and keep their commit points intact when the input gets strange.
PILLAR 05The coordination layer underneath
Notice what the recruiting agent, the SDR agent, the field-service agent, and the clinical outreach agent all have in common. Each one, after it does its specialized work, has to commit to a time. A demo. An interview. A service window. A follow-up call. Coordination is not a feature of one vertical. It is the connective tissue the entire agent economy runs on.
The numbers point the same direction. More than half of talent leaders say they plan to add autonomous agents in 2026, and the leading recruiting agents already describe their product as a source-to-schedule loop, where finding the candidate and booking the conversation are one continuous motion rather than two disconnected steps. The same shape shows up in sales, in field service, and in clinical outreach. The specialized work and the scheduling are not separate phases. The scheduling is the moment the work turns into a commitment.
That is the quiet thesis behind the whole list. The agents that capture attention are the ones doing visible, specialized work. The agent that captures value over the long run is the one every other agent has to call to get anything onto a calendar. In the agent era, scheduling stops being a utility and becomes infrastructure.
Consider the mechanics. A recruiting agent can source and screen a hundred candidates on its own, but the instant it needs to book a panel interview across four busy calendars in three time zones, it hits a coordination problem that has nothing to do with recruiting and everything to do with negotiation, constraints, and trust. Every vertical agent eventually arrives at that same wall. Whoever owns the layer that resolves it owns a position that compounds as the number of agents in the economy grows.
From AI Edge for Leaders
Read the architecture, not the demo
The ten agents above are real, deployed, and already earning their keep across software, support, sales, law, medicine, field service, finance, security, hiring, and coordination. What unites them is not the model underneath, which is fast becoming a commodity. It is agent-first architecture, codified procedure that took years to build, and a commit point placed exactly at the line between reversible and irreversible action. For leaders, the mandate is simple to state and hard to fake. Judge agents by their work product, not their demo. Insist on knowing where the human stays in the loop. Then build to that standard, or buy to it.
Read the full issue