P01 · Agent Runtime
Agent engineering with explicit tool boundaries: database access, reusable AI assets, browser operations, command execution, file editing and persistent task context.
Agent execution and state ownership belong behind an explicit runtime boundary—not inside UI events or API handlers.
Agent execution is treated as an explicit runtime boundary, rather than logic scattered across UI components or API handlers.
Further notes on execution lifecycle, state ownership and integration boundaries are being organized. This page currently presents the core design principle rather than a complete runtime specification.
Public project evidence and purpose-built development tools support governed Agent integrations, from database access to durable project workflows.
A database MCP gateway with a public source repository, and a live platform for reusable Agent assets.
A MySQL/PostgreSQL MCP gateway with scoped endpoints and token-based user ACL. The Connection → Endpoint → ACL model supports table whitelists; SQL-AST read-only checks, audit logs and usage metering constrain database access.
An AI work-asset platform for managing prompts, skills, project context, tool specifications and versioned resources. It turns reusable Agent context into organized, discoverable and access-governed assets.
Browser observation, command execution, file operations and persistent planning each have a separate operational boundary.
Inspect real pages and interactions.
Open tabs, click elements, type, capture screenshots and inspect DOM, network events and console output.
EXAMPLE · Reproduce a UI failure or trace an AI SDK initialization sequence. Network events are not a complete HAR or guaranteed response-body capture.
open_urllist_tabsclick_snapshottype_textget_network_eventsget_console_logscapture_screenshotexport_storage
Run controlled commands and persistent sessions.
Choose direct argv execution, one-shot shell commands or interactive sessions; inspect output, continue input and terminate processes.
EXAMPLE · Run Python checks, tests and builds; keep a development server running during browser verification.
exec_task_createexec_shell_task_createexec_session_create
Read, modify and validate workspace files.
List directories, batch-read source files, inspect log ranges, save files and validate or apply small unified patches.
EXAMPLE · Use save_file for large edits; use patches for localized changes. Tool output returns file content—not parsed JSON semantics.
batch_list_directoryread_filessave_filevalidate_unified_patchapply_unified_patchtail_file
Carry decisions and evidence across sessions.
Persist long-running objectives, work items, dependencies, risks, milestones and validation evidence. Resume work and export handoffs.[1]
EXAMPLE · A gad-exec task runs one program; a gad-plan task tracks a long-lived engineering goal.
task_starttask_nexttask_commit_phasetask_linttask_export_handofftask_export_timeline
Tools perform actions. Project state retains decisions, constraints and verification across model and session boundaries.
gad-plan establishes a durable task and next steps.
gad-browser observes the UI, requests and console.
gad-exec executes tests, builds and services.
gad-filesystem reads, saves or patches files.
gad-plan records evidence, decisions and handoff.
DDD and TDD help keep state ownership, interface contracts and regressions explicit as the Agent system evolves. Tool execution is separate from verification: tests and evidence determine whether a change is accepted.
Direct tool documentation comes first. Other articles offer context on the engineering workflow or enterprise applications; they are not presented as proof of KDD implementation details.
The engineering note associated with the persistent project-planning tool.
https://pylab.me/zh-CN/posts/gad-plan-agentic-project-infrastructure/Related reading on Codex-oriented prompting and working practices.
https://pylab.me/zh-CN/posts/best-practices-intj-prompts-for-codex/A related perspective on working with AI.
https://pylab.me/zh-CN/posts/thoughts-on-using-ai/A separate enterprise AI application example, not evidence for KDD runtime internals.
https://pylab.me/zh-CN/posts/audit-to-algo-dify-ai-practice/Ask about MCP server boundaries, database permissions, browser debugging or persistent agent-development workflows.