From Copilots to Coworkers: The Rise of Agentic Workflows and the Infra They Demand

Thesis — agents are graduating from suggestions to action

The last two years looked like a parade of copilots: models that help you write, search, or summarize. The next wave — which the YC cohort above is quietly building — is autonomous and semi-autonomous agents that complete end-to-end work across sales, support, coding, and operations. These are not UI toys. They are systems that make decisions, call APIs, and change state in production. That shift converts a UX problem into an infrastructure problem: identity, orchestration, observability, security, data, and governance become first-order product levers.

The landscape — four practical sub-patterns

Below I group the companies by the kind of work their agents are doing and the infra they imply. Each group shows how agent capabilities are being packaged and sold — and where the hard engineering lives.

Horizontal orchestration and developer primitives Agents need ways to act: connect APIs, run code, schedule tasks, and ship containers. That’s where platforms that focus on orchestration and execution live. Zapier converts integrations into agentable automations; Docker remains the substrate for running autonomous workloads; Apollo is the API glue for agentic flows; GitLab is turning its DevSecOps suite into an agent orchestration and automation layer.

S12

Zapier★

The easiest way to automate your work.

ActiveB2B700 ppl

S10

Docker

Software development platform.

ActiveB2B201 ppl

S11

Apollo

Apollo helps engineering teams accelerate delivery

ActiveB2B200 ppl

W15

GitLab★

A complete DevOps platform delivered as a single application.

PublicB2B2000 ppl

Customer-facing agents — voice, chat, and CX automation The most visible agents are customer-facing: they take calls, triage tickets, and handle conversations. Observe.AI builds purpose-built agents for contact centers; Plivo supplies voice and messaging agents with ASR/TTS and a no-code studio; Sendbird is embedding concierge-style agents into messaging and voice; Jasper.ai runs marketing agent pipelines that execute content ops end-to-end.

W18

Observe.AI

Contact center AI platform.

ActiveB2B240 ppl

S12

Plivo

Voice AI Agents for customer engagement, including WhatsApp, SMS &…

ActiveB2B320 ppl

W16

Sendbird

The AI agent that doesn’t just support, it delights.

ActiveB2B300 ppl

W18

Jasper.ai

Jasper is a leader in creating content using Artificial Intelligence

ActiveB2B195 ppl

Decisioning and regulated workflows Where decisions carry risk — credit, AML, compliance — teams prefer human-in-the-loop, auditable agents. Taktile and Zip are building agentic decision platforms for finance and procurement. Simetrik automates reconciliation (an agentic task), and Middesk supplies the KYB signals that let agents decide to onboard or block a counterparty. Klarity maps organizational process context so agents can act without breaking established controls.

S20

Taktile

Transform your decision-making with reliable AI agents in weeks, not…

ActiveB2B150 ppl

S20

Zip

The future of B2B spend

ActiveB2B500 ppl

W18

Simetrik

No code enterprise SaaS for financial control automation.

ActiveB2B350 ppl

W19

Middesk

We make it easier for B2B companies to transact with one another

ActiveB2B130 ppl

Observation, identity and security — the agent control plane Agents change state; you need telemetry, identity, and policy. Mezmo positions an agent control plane for logs and incident remediation. Encord is the data layer for physical AI (training agents that operate in the real world). Teleport treats humans, machines, and AI as identities to be governed. Salt Security maps the "Agentic Security Graph" — a neat phrase that captures how API/agent sprawl creates new attack surfaces.

S15

Teleport

The Infrastructure Identity Company

ActiveB2B255 ppl

W15

Mezmo

Observability platform to control all of your log data

ActiveB2B172 ppl

W21

Encord

The data layer for physical AI

ActiveB2B150 ppl

W16

Salt Security

Protects organizations from getting breached through their APIs.

ActiveB2B200 ppl

Physical and industrial agents Autonomy in the real world is a different economy: robotics and autonomous vehicles embed perception, controls, and long-running agentic decision systems. Bear Flag Robotics (acquired), Ginkgo Bioworks, Embark Trucks and May Mobility show the investment and exit paths when agents replace manual labor or lab bench work.

W18

Bear Flag Robotics★

Autonomous driving technology for farm tractors.

AcquiredIndustrials10 ppl

S14

Ginkgo Bioworks★

Our mission is to make biology easier to engineer.

PublicHealthcare641 ppl

W16

Embark Trucks

Self-driving semi trucks.

PublicIndustrials300 ppl

S17

May Mobility

Transforming cities and rural areas through AV transit and technology

ActiveIndustrials260 ppl

Batch cohorts — the concentrated bets

Cohorts tell us what accelerators were *thinking* at a moment in time.

• Winter 2016 is a clear early cluster: Luciq, Salt Security, Sendbird, Relativity Space, and Embark Trucks — a batch that mixed observability, security, and physical autonomy, suggesting YC saw agentic control planes and robotics as adjacent plays. That adjacency matters: a lot of the agent stack reuses patterns (telemetry + decision loop + safety).
• Winter 2018 groups another concentrated bet on enterprise agents and governance: Vanta, Observe.AI, Jasper.ai, and Simetrik. This cohort signals a batch-level focus on operationalizing AI inside compliance-heavy businesses.
• Summer 2020 shows verticalization: Zip, Taktile, and Fieldguide — agents specialized for procurement, credit/AML, and professional services.

Those adjacent batches (early infra in 2012–2015 with Rescale, Algolia, GitLab, Mezmo) form the scaffolding: cheap compute, retrieval/search, and unified CI/CD made agentic workflows feasible.

Shared characteristics

Across verticals the winning trade-offs repeat:

• Integration-first: agents act by connecting existing systems (Zapier, Apollo, Middesk). Buyers choose the agent that plugs into their stack.
• Programmability + No-code: developer APIs run the hardest paths; no-code studios (Plivo, Jasper.ai) bring non-technical operators into play.
• Human-in-the-loop safety: regulated buyers demand audit trails and decision overrides (Taktile, Simetrik, Vanta).
• Infra-first bets: telemetry, identity, and policy are commercial primitives — not optional add-ons (Teleport, Mezmo, Salt Security).

What’s working — exits, public names, and pivots

Public and acquired outcomes validate the category: GitLab, PagerDuty, and Ginkgo Bioworks are public — proof that orchestration and agentic automation can scale commercially. Bear Flag Robotics was acquired (industrial autonomy is M&A-friendly). Many agentic companies show adaptive product strategy — several list PIVOT: yes (renamed) — Sendbird, Luciq, Mezmo, and Simetrik among them — which reads as iterative moves from discovery to action.

Risks and tarpits

• Integration fragility: agents are only as useful as the signals and actuation points they can reliably access. The long tail of enterprise systems is a maintenance tax.
• Security and identity: agents multiply privileged access. If you don’t solve machine and AI identity first, you create a brittle, exploitable surface — which is precisely the space Teleport and Salt Security are addressing.
• Observability vacuum: autonomous actions require explainability and fast RCA; that’s the core friction Mezmo and Luciq are trying to close.
• Economics and governance: agents shift labor spend into compute, data, and oversight costs. Vendors that ignore procurement and compliance (Zip, Vanta, Taktile do not) will lose enterprise deals.

Why now — and the near-term outlook

LLMs made agent behaviors natural; cheap cloud compute, mature container tooling, searchable retrieval (Algolia), and orchestration primitives (Docker, GitLab, Zapier) make them practical. The next 18–36 months will be consolidation around four rails: identity (Teleport), orchestration (Zapier/GitLab/Apollo), observability (Mezmo/Luciq/Encord), and policy/security (Salt Security/Vanta). Vertical leaders will win where domain risk is highest — finance, healthcare, logistics — because those buyers will pay for audited, reversible automation.

If you’re an investor or operator: bet on infra that makes agents safe and observable, and on vertical agents that can prove ROI in a quarter. The rest will be a story of integrations and governance — boring, painful, valuable work that turns copilots into coworkers.