Experiment
Start in interactive mode. Private by default: the run and its files stay with whoever created them.
From one-off experiment to operational workflow.
The platform organizes agent work around a maturity cycle: experiment, observe, review, repeat, govern and automate. Every run leaves evidence; autonomy is earned over time.
The maturity cycle of work
Autonomy is earned, not granted. The cycle defines when a person steps in, what is recorded and what can be reused.
Observe
Every run records time, cost, agent and model, events and permitted access.
Review
A person assesses the result and the recorded decisions before the work is used.
Repeat
Re-run the same workflow with new inputs and compare behavior across runs.
Govern
Define access, limits and network policies for the flow. Sharing is explicit.
Automate
When the evidence shows maturity, reduce human intervention and increase autonomy.
Privacy, review and autonomy at every stage
Privacy
Runs and files start private to their creator; sharing is explicit.
Human review
The result of every run goes through a person. Approval during execution still needs a specific demonstration.
Policies
Access, limits and network destinations are defined before the run and recorded throughout the cycle.
Evidence
Inputs, output files, events and decisions are available for inspection on every run.
Progressive autonomy
From interactive mode to runs with less intervention, each step builds on accumulated evidence.
Customer-controlled persistence
History and evidence can be persisted in customer-controlled AWS resources.
Reuse resources from the catalog
The catalog supports the prepare and repeat stages of the work. Choose what the task needs and review instructions, versions and access before running.
Prompt templates
Reusable instructions and inputs for recurring tasks.
Skills
Specialized instructions that guide the agent through a task.
Base images
Prepared environments with defined versions of coding agents and their tools.
Network policies
Rules that specify the permitted network destinations for a run.
MCP connectors (Model Context Protocol)
Connections configured with tools and information sources. Review the required accesses before using each connection.
Agent packages
Sets of instructions and files ready to configure an agent. Check package requirements before choosing the environment.
Use context from your GitHub repository
Specify the repository, the desired version and an authorized access connection. The GitHub input workflow prepares a copy of the repository before starting the agent. In the demonstration, inspect which files were made available for the task.
Approval during execution: part of the design
Approval while the agent works still needs a specific demonstration. It is distinct from reviewing the final result. The pilot scope defines how to assess this requirement in your flow.
See a run and its records
In the 30-minute meeting, we walk through the maturity cycle and a live or previously prepared run. You see the inputs, the output and the available records. For a GitHub scenario, we inspect repository context. We confirm the scenario when scheduling and define what to assess in the pilot.