Summary: The agent market got noisier in 2026 but easier to read, because every vendor now fits one of five categories: visual builders, code frameworks, hyperscaler infrastructure, forward-deployed services, and compiled-execution platforms. The tell is not the demo. It is who builds the workflow, who owns the logic, and what actually happens at runtime. Here is the map as of July 2026, including what the summer’s frontier-model barrage did and did not change.
Five categories, three questions
Strip the branding from any agent vendor and three questions place them on the map:
- Who builds the workflow? Your team on a canvas, your engineers in a framework, their engineers on an engagement, or a compiler working from your policies.
- Who owns the logic? Can you read it, export it, and take it with you, or does it live in the vendor’s proprietary format?
- What happens at runtime? Reasoning loops that improvise per case, or a deterministic plan that executes the same way every time.
Category 1: Visual builders wearing agent costumes
Microsoft Copilot Studio, Power Automate, UiPath Agent Builder, Workato, Automation Anywhere, Salesforce Agentforce. The lineage runs straight back to the 2011 recipe-and-flowchart paradigm, now with an LLM inside each box. Strong ecosystems, real productivity wins on simple flows, and the same structural ceiling they have always had: every edge case is another branch, every rule change is a re-wiring session, and the diagram drifts from production. Our comparison library covers each in depth.
Category 2: Code frameworks
LangChain and LangGraph, CrewAI, AutoGen, Semantic Kernel. The most flexible category and the most expensive to operate: you own orchestration, retries, evals, observability, and compliance infrastructure forever. Prototypes take hours; production regulated workflows take 6 to 18 months. Frameworks are the right answer when the agent platform is itself your product. They are a commitment, not a shortcut.
Category 3: Hyperscaler infrastructure
Amazon Bedrock, Google’s Gemini Enterprise umbrella over Vertex AI, OpenAI’s Frontier and Workspace Agents. World-class primitives, priced attractively, with the assembly problem left to you: document pipelines, policy enforcement, and audit-grade traceability are integration projects on top. The summer’s launches deepened the primitives without closing that gap.
Category 4: Forward-deployed services
The newest category to reach scale, and Poetic is its flagship: operators describe procedures, forward-deployed engineers compile them into a proprietary language, and the vendor runs the automation for you. Real results at marquee logos, with a structural trade: you get outcomes, but the vendor owns and maintains the artifact. Palantir’s AIP-plus-forward-deployed model rhymes with this category from the data-platform side.
Category 5: Compiled-execution platforms
The category MightyBot occupies: describe the agent in plain English, upload the governing policies and documents, and the platform compiles the schemas, workflow, and execution plan. The logic stays yours, in human-readable versioned files. Runtime is deterministic, which is what makes 99%+ production accuracy and evidence-linked audit trails achievable at volume in regulated operations.
What the summer model wars changed
Between June 13 and July 9, GLM 5.2, GPT-5.6 Sol, Terra, and Luna, Grok 4.5 with Grok Build, and the Claude 5 family all shipped. Prices moved down; benchmarks reshuffled twice.
For buyers the lesson is the opposite of the hype: model capability is now a rapidly depreciating differentiator. Whatever model advantage a vendor claims today erodes in a quarter. The properties that persist are architectural: who builds, who owns, what runs. A model-neutral platform banks every model improvement automatically; a platform welded to one lab’s models makes you renegotiate the stack every launch season.
Using the map
Ask the three questions in every evaluation. If the answer to “who builds the workflow” is a canvas, budget for the maintenance of the diagram. If it is “our engineers, per engagement,” read the terms for who owns the result. If it is “a compiler, from your policies,” check that the execution is genuinely deterministic and the audit trail is generated, not reconstructed. Then price all of it: the ROI calculator models the build paths against the platform path with measured token data.
