Chatbots answered questions. Copilots helped with discrete tasks. The next layer is agentic systems: software that can plan multi-step work, use tools, maintain state, and operate with limited supervision toward a goal. The demos have been impressive for some time. The products are beginning to arrive.
What “agentic” actually means in practice
In current systems, an agent typically combines several ingredients:
- A planning or reasoning loop that breaks goals into steps
- Access to tools (browsers, code interpreters, APIs, internal systems)
- Memory or state that persists across steps and sessions
- Mechanisms for verification, error recovery, and human escalation
Reliability is still the binding constraint. Agents fail in ways that are hard to predict: they loop, they invent intermediate facts, they misuse tools, or they stop short of the goal. The startups that matter will be the ones that turn partial reliability into useful, bounded products rather than open-ended general assistants.
Where early products are forming
Several categories are taking shape:
Research and knowledge work agents
Systems that can survey literature, summarize sources, draft analyses, and keep living documents up to date. These work best when the domain is text-heavy and the cost of an occasional error is manageable.
Software engineering agents
Agents that can explore a codebase, propose changes, run tests, and open pull requests. The feedback loop of tests and type systems makes this domain more tractable than many others.
Operations and workflow agents
Agents that handle repetitive multi-step processes across SaaS tools — ticket triage, report generation, data cleanup, onboarding sequences. Success here depends on deep integration and clear success criteria.
Domain-specialized agents
Legal review, clinical documentation support, financial analysis, and similar vertical applications. Specialization reduces the action space and raises the bar for quality, which can make agent behavior more predictable.
The startup opportunity
Building a general-purpose agent that works everywhere remains extremely hard. Building a reliable agent for a narrow, high-value workflow is achievable. That is where most early commercial success is likely to appear.
Defensibility will come less from the base model and more from:
- Proprietary evaluation data and feedback loops
- Deep tool integrations and permission models
- Domain-specific scaffolding and guardrails
- Trust and compliance surfaces that enterprises require
Infrastructure startups are also forming around agent needs: better observability for multi-step runs, safer tool execution environments, memory and retrieval systems designed for long-horizon tasks, and evaluation platforms that can score agent trajectories rather than single responses.
Research that feeds the product layer
Progress on planning, tool use, multi-agent coordination, and verification is still heavily research-driven. Open experiments and small-lab prototypes continue to surface ideas that larger product teams later adopt. Efforts such as the AI research project at labs.maxpri.workers.dev sit in that intermediate zone — exploring capabilities that may later become product primitives.
What to watch
Over the next 12–18 months, the important signals will be:
- Whether any vertical agent products achieve retention and willingness-to-pay that looks like durable software businesses.
- Whether evaluation methods improve enough that teams can ship agents with measurable reliability guarantees.
- Whether multi-agent systems move beyond demos into coordinated real workflows.
Agentic systems will not replace every form of software. They will create a new layer on top of existing tools and data — and the companies that own the most useful, trustworthy instances of that layer will define a new category.