An agent is not a chatbot

The word agent got used for everything in 2025. In 2026 the industry has settled on a workable definition: an AI agent is software that receives a goal, plans the steps to reach it, calls tools or other agents to execute those steps, and reports back a finished result. A chatbot answers a question. An agent finishes a task. The difference matters — because everything about how you evaluate, buy, deploy and govern one is different from a chatbot.

IDC projects agent use will rise by an order of magnitude in the next few years, with hundreds of millions — and eventually a billion-plus — operating inside enterprises. That is not marketing weight; it is the reason Microsoft, OpenAI, Anthropic, Google and Salesforce all launched enterprise agent platforms between March and May of 2026. This piece is the honest map: what agents genuinely do in production today, where they still break, the 2026 platform landscape, and how to start.

The 2026 platform landscape

Five serious enterprise platforms, all released or repositioned in the first half of 2026. OpenAI Workspace Agents (April 22, 2026) replaced Custom GPTs with cloud-executed agents running on the Codex model, aimed at team collaboration and asynchronous work. Claude Agent SDK from Anthropic (enterprise pricing from $50,000/year, 50 seats minimum) leans into deep reasoning on long context (up to a million tokens on Opus 4.8), local execution, and native MCP integration. Microsoft Copilot Studio + Agent 365 (Agent 365 generally available May 1, 2026 as part of Microsoft 365 E7 at $99/user/month) is the most complete governance-first stack — low-code Copilot Studio, computer-using agents that drive legacy apps through the screen, and Agent 365 as the control plane. Google Gemini Enterprise Agent Platform (launched the same day as Workspace Agents) unifies Vertex AI's model selection, model building and agent building, with A2A (Agent-to-Agent) protocol support and an interesting Agent Memory Bank for long-term memories. Salesforce Agentforce 360 is the incumbent play: sitting on the Salesforce data cloud, deep CRM integration, right for anyone already living in the ecosystem.

The convergent pattern across all five: multi-agent orchestration. One agent no longer does everything — a supervisor agent delegates to specialist agents (research, code, data, ops), each with its own tools and model. Microsoft's Agent Framework 1.0 (Semantic Kernel + AutoGen merged) is the most complete open-source stack; the Anthropic SDK and OpenAI Agents SDK cover the same pattern with different ergonomics.

Where agents genuinely deliver in 2026

1. Long, multi-step research and writing. Copilot Cowork (Claude-powered inside Microsoft 365) runs multi-day workflows: pull data from four systems, draft the deck, iterate on tone, extract action items. Anthropic's Claude Code has a “dynamic workflows” feature orchestrating hundreds of parallel sub-agents for large-scale codebase migrations.

2. Tier-1 customer support with human escalation. Agents resolve the 60-70% of tickets that are password resets, order status, refund requests and simple product questions. The remainder route to humans with full context attached. The economics work at scale — not because the agent is smart, but because the routing and context handoff finally does.

3. Legacy application automation. Microsoft's computer-using agents drive desktop and web applications through the screen — clicking, typing, reading the UI — where no API exists. This is the RPA replacement enterprises have been trying to build for a decade; it now genuinely works for constrained workflows.

4. Sales and CRM data entry. Salesforce Agentforce 360 finally answers the question every VP of Sales has asked for twenty years: why do the reps not update the CRM? Answer: because the agent does it from meeting recordings, calendar entries and email threads. Adoption is measurable in weeks, not quarters.

5. Engineering and code migration. Claude Code’s dynamic workflows and OpenAI Codex's parallel execution both handle multi-repo refactors, dependency upgrades and legacy translation at scale. The pattern is the same as the others: agents don’t replace engineers, they finish the tedious 60% so engineers can spend time on the interesting 40%.

Where they still break

Prompt injection remains real. Capsule Security's CVE-2026-21520 exposed 200,000 MCP servers to remote code execution risk before Microsoft patched Copilot Studio in January 2026. Every agent that touches live systems and takes actions across previously separated application boundaries carries this risk at elevated consequence. Governance layers (Agent 365, enterprise MCP scanning) do not eliminate the risk — they make it observable and manageable.

Long-context reliability drops. Multi-day agent runs still degrade in tricky ways: memory drift, tool-call retries that duplicate side effects, decisions that made sense at step 4 but not at step 40. The mitigation is architectural — checkpoint memories, idempotent tools, human review gates — not just a smarter model.

Cost surprises are frequent. Claude Opus 4.8 is token-inefficient on sophisticated outputs; a dynamic workflow parallelising hundreds of sub-agents can burn through API credits in an afternoon. Copilot Studio’s 25,000 Copilot Credits at $200/month look generous until an agent enters a loop it shouldn’t. Budget alerts and per-agent cost ceilings are as important as functionality.

Governance is not optional. Microsoft’s Agent 365 — the reason E7 costs $99/user/month — exists because enterprises need audit trails, permission scopes, kill switches and shadow-agent discovery. Any deployment without governance will produce a compliance incident before it produces measurable value.

The right way for an enterprise to start

The mistake we see is starting from the platform (“let’s do Copilot Studio”) instead of the use case. The pattern working teams use in 2026 is inverted. Pick one narrowly scoped, tedious, observable workflow — something like “triage inbound support tickets and draft replies with product context.” Build it on the platform where your existing stack lives (Microsoft 365 shop → Copilot Studio; Salesforce shop → Agentforce; deeply custom → Claude Agent SDK or Anthropic API direct). Measure two numbers weekly: time saved per resolved case, and rate of human overrides. If both trend the right way for three months, expand scope. If they don’t, kill the pilot and try a different use case — not a different platform. Platforms are 20% of success; the use case is 80%.

Two things day one: someone on the team owns the agent (not IT, not a vendor), and a governance layer exists (Agent 365, or an equivalent from the platform you picked). Without both, you are building the compliance incident, not the productivity multiplier.

What we build at DZDSoft

DZDSoft integrates AI agents into production workflows for clients across construction, façade engineering and enterprise services. What we don’t do is treat agent deployment as a magic-button project. We start with the tedious workflow that has clear success metrics, we pick the platform that fits the existing stack (not the loudest one), and we ship the governance layer before the first agent goes live. If you want to explore what an agent could do for a specific workflow in your business — and hear an honest read on whether it’s the right use case — that’s the conversation we have with clients every week. Bring the workflow. We’ll bring the honest answer.