Clear Ideas Agent Runtime
August introduces Clear Ideas Agent Runtime, a standalone, provider-neutral TypeScript runtime for defining, running, and observing portable AI agents.
Agent Runtime can run locally, inside an application, in a child process, or on remote compute. It does not require a hosted Clear Ideas service. Developers can install the runtime and CLI from npm without building the project from source, choose the model and infrastructure adapters they need, and operate the runtime independently under the Apache License 2.0.
What's new
- Portable Agent Manifests for defining steps, variables, conditions, loops, tools, outputs, and limits as versioned YAML or TypeScript contracts
- Separate Agent Run Manifests for supplying validated run-time values without changing the underlying agent definition
- Ready-to-install npm packages and CLI for using Agent Runtime without a source build
- Native authorization boundaries for credential resolution, connections, tool allowlists, access modes, and host authorization decisions
- Native sandbox and artifact contracts with included Docker and Modal providers and interfaces for custom infrastructure
- Durable execution state with checkpoints, suspension, cancellation, fresh-process resume, attempt fencing, and ordered events
- Manifest graph execution that resolves variable dependencies into a safe fan-out and fan-in schedule with manifest-order commits
- Local and remote execution options through in-process, child-process, worker-protocol, and remote execution engines
- Provider-neutral adapters for models, persistence, compute, sandboxes, generated files, and OpenTelemetry
- Developer documentation and examples covering manifests, embedding, streaming, recovery, connections, custom adapters, and production operation
Why it matters
An agent definition should not have to be rewritten when it moves from a local prototype into an application or remote execution environment.
Agent Runtime separates the portable agent contract from credentials, models, compute, storage, sandboxes, and other host-controlled infrastructure. That makes an agent easier to inspect, test, move, and operate while keeping security-sensitive decisions outside the manifest itself.
The project is a standalone agent runtime, not a hosted-service requirement. Clear Ideas uses it as an execution foundation, but developers can install and run it independently in their own applications and infrastructure.
Agent Runtime Foundation for Clear Ideas Agents
Clear Ideas has also integrated the new runtime foundation into governed agent execution.
This brings the portable execution contracts, durable run state, ordered events, adapter boundaries, and recovery behavior from Agent Runtime into Clear Ideas Agents while retaining the organization policies, governed records, content permissions, and managed operations of the Clear Ideas product.
What's new
- A shared portable execution foundation across local and remote agent environments
- More durable run handling for prompts, loops, approvals, webhooks, code steps, sub-runs, and generated files
- Stronger continuation and cancellation behavior for longer-running or interrupted agent processes
- Consistent artifact and event projection into Clear Ideas Agent Run records
- Improved terminal-state reconciliation when a remote worker completes, fails, or is replaced
- Expanded remote-run observability for execution health, timing, generated files, and infrastructure capacity
Why it matters
Governed agent processes often run longer than a normal chat response and may cross several execution boundaries. They need to preserve state, recover predictably, and leave a useful record even when a worker restarts or a run is resumed elsewhere.
The Agent Runtime integration gives Clear Ideas Agents a more consistent execution model while preserving the product-level controls that organizations use to govern models, content, users, tools, and records.
Manifest Graph Execution
Agent Runtime can resolve a declarative Agent Manifest into a dependency-aware execution plan.
Sequential execution remains the default. When a run enables graph scheduling, Agent Runtime reads step outputs and variable dependencies, groups eligible work into execution waves, fans independent prompt branches out concurrently, and fans them back in before dependent work continues. Results commit in the order defined by the Agent Manifest.
What's new
- Manifest-to-graph resolution without a separate graph definition or infrastructure-specific rewrite
- Dependency-aware fan-out and fan-in for independent steps that can safely run at the same time
- Live streamed output by step while eligible parallel work is in progress
- Manifest-order result commits so completed outputs remain predictable
- Ordered execution for stateful work including dependent steps, loops, and tool calls
Why it matters
The runtime derives the execution graph from the portable manifest. Parallelism is one result of that plan: independent work can proceed concurrently while orchestration and state commits remain ordered.
Agents that perform independent research, classification, extraction, or drafting before synthesis can complete substantially faster, depending on graph shape, model latency, and provider limits. The option remains explicit and scoped to eligible work, keeping side effects and stateful operations ordered.
PowerPoint Previews and Richer Presentation Generation
Generated presentations are easier to review directly inside AI Chat.
Users can now preview generated PowerPoint files before downloading them. Presentation preview rendering is cached and coordinated so repeated previews are faster and multiple requests do not produce unnecessary duplicate work.
What's new
- In-chat previews for generated PowerPoint files
- Clear preview and download actions for presentation artifacts
- Cached preview rendering for faster repeat access
- Improved handling for multiple presentation outputs
- A richer generative presentation skill for more art-directed layouts and visual composition
- Better presentation streaming and status handling while a deck is being generated
- English and French presentation-preview messages
Why it matters
A generated presentation should be easy to inspect before it leaves the governed AI experience.
Previewing the deck in context helps users confirm its structure and visual direction before downloading the .pptx file, while richer generation gives the presentation system more flexibility than a fixed slide template.
Model Catalog and Reasoning Controls
Clear Ideas now presents model and reasoning options with more model-specific detail.
The model catalog has been updated so reasoning controls reflect the options actually supported by the selected model. Administrators also have clearer model metadata and access controls, while users see a cleaner selector that avoids presenting unsupported reasoning choices.
What's new
- Model-specific reasoning options instead of one common list for every reasoning model
- Additional reasoning levels where the selected model supports them
- Clearer reasoning information in model selectors and tooltips
- New Cerebras model integrations for approved AI experiences
- An expanded generative-presentation model path
- Updated AI capability information, including newer provider and model entries such as Grok 4.5
- Refined organization model controls for visibility, availability, and policy-aware selection
Why it matters
Model capabilities are not interchangeable. Reasoning levels, structured output, visual input, context size, speed, and provider handling can vary substantially.
These changes make model selection more accurate for users and easier to govern for administrators, while allowing Clear Ideas to add new model paths without presenting controls that a particular model cannot use.
Supporting Improvements
August also includes ongoing reliability, security, and operational work across Clear Ideas.
- Governance-aware search filtering now more consistently applies AI eligibility and indexing exclusions to content results
- Improved content synchronization for renamed, deleted, and tombstoned folders and files
- More consistent content selection behavior across list and tile views
- Stronger profile and account synchronization with less unnecessary background traffic
- Improved email verification and Google Sign-In return handling
- Better API request timing, health checks, and operational diagnostics
- Expanded backup and restore monitoring
- Additional remote-runner health and capacity metrics
- Dependency and security maintenance across the web application, API, and deployment infrastructure
Together, these changes make governed search, account access, agent execution, and core platform operations more dependable.