Some developments change which model you choose. Others change how you build the whole workflow. We separate shipped capabilities, early access and provider claims.
Our question is simple: what useful work becomes possible, and what still needs to be proven? These are researched explainers with editorial analysis, not hands-on performance tests.
TypeSafe AI · 15 Sept 2026Early access; waitlist
Jev / System One Models
Typed decisions with explicit uncertainty
A specialised model for choosing, scoring and routing inside software. Jev is not a general chatbot or a replacement for a reasoning agent.
What changes
Developers send state plus narrowly defined questions. Choice returns an option and probabilities; Score returns a rubric score and probabilities; Noul returns a 0–1 assessment. Choice and Score also include confidence derived from their distributions. Questions run in parallel. TypeSafe calls its training method Reinforcement Learning for Calibrated Decisions (RLCD); ordinary code combines the decisions.
Why it matters at work
Many agent workflows need repeated small judgments more often than open-ended prose. A constrained decision component could make these branches cheaper, faster and easier to review.
+ The promise
Typed values can drive code without parsing generated prose.
Probability distributions let developers route ambiguous cases.
Parallel evaluation can reduce latency for independent questions.
− What to question
Cannot generate arbitrary text; complex reasoning must be decomposed or delegated.
Schema validity does not guarantee a factually correct decision.
Speed, cost and calibration results are vendor-reported and workload-dependent.
The human role: People define the options, rubrics and escalation rules, validate calibration on their data, and review uncertain or consequential cases.
Customer request→Jev: classify & score→Your policy & permissions→Agent action or human review
Illustrative architecture. Valid types do not guarantee a correct decision, and probability thresholds need testing on your own cases.
AnthropicBeta; enabled by default for API accounts
Claude Managed Agents
Durable sessions become a managed service
A hosted agent harness handles execution and persistent sessions around Claude. This is infrastructure, rather than a new foundation model.
What changes
The service supplies the agent loop, sandbox, tools, caching and context compaction. Long-running sessions retain conversation history, sandbox state and outputs, and can resume after pauses. Built-in tools cover files, shell and web access; MCP connects external tools.
Why it matters at work
A model's context window alone cannot keep a project running. A managed runtime reduces the engineering needed to resume work and coordinate execution.
+ The promise
Persistent sessions preserve work across pauses.
Built-in execution tools reduce integration work.
Caching and compaction are handled by the harness.
− What to question
Beta behavior may change.
State and outputs are stored server-side.
Currently ineligible for Zero Data Retention and HIPAA BAA coverage.
The human role: People set the objective, configure tools and access, check progress and validate outputs. Managed execution does not replace domain accountability.
Model Context Protocol maintainers · 26 Jan 2026Stable official extension
MCP Apps
Interactive tools inside the agent conversation
MCP tools can return forms, dashboards and other interactive interfaces, giving people a direct way to inspect and steer agent work.
What changes
A tool declares a UI resource. A supporting host renders the interface in a sandboxed frame, and the UI communicates with the host through the extension's protocol. This is an interoperability layer for applications, not an AI model.
Why it matters at work
Reviewing a chart or adjusting a form can communicate decisions more precisely than another paragraph of chat, especially when a workflow needs human approval.
People can inspect and adjust structured information directly.
Interactive views stay alongside the task conversation.
− What to question
The host must support the extension and its required features.
Developers still build and maintain the interface and backend.
An embedded UI does not itself guarantee correct data or safe actions.
The human role: People use the controls to inspect results, change parameters and approve intended actions; applications still define authorization.
GooglePreview capability on a stable model
Gemini 3.8 Flash Computer Use
Screen interaction with explicit action intent
An agent can inspect screenshots and propose clicks, typing and navigation across browser, desktop and mobile environments.
What changes
The model receives the task and screen state, returns an action with its intent, and the application executes it and sends back updated state. Gemini 3.x can return safety decisions requiring confirmation. Prompt-injection detection is available but opt-in.
Why it matters at work
Screen-based tools extend automation to software without a suitable API. They also make the observation–action–verification loop visible.
+ The promise
Reaches workflows that expose only a graphical interface.
Action intent helps people understand the next step.
Custom tools can explicitly yield control to a person.
− What to question
The computer-use capability remains a preview.
UI mistakes and security vulnerabilities remain possible.
The application must implement execution, confirmation handling and recovery.
The human role: People supervise important tasks, handle requested confirmations and validate the result. Use recoverable workflows while evaluating reliability.
A CLEARER VIEW OF WHAT COMES NEXT
A lot is changing. Stay a step ahead.
Your briefing on the ideas, agents and robots reshaping work.