Running an agent is straightforward compared with operating it. A team also needs to see the task state, understand which tools are available, review sensitive actions, and trace the result later. Separate dashboards and scripts make that context difficult to follow.
This project explores a shared operational foundation: tasks, sessions, policies, integrations, and audit records connected through one workspace.
02 / Architecture
Separate responsibilities. Preserve the record.
The dashboard presents operations. The runtime manages task execution and sessions. The policy layer handles checks and approvals. The MCP hub connects tools. PostgreSQL and pgvector provide persistent state and memory.
Conceptual component map based on the repository architecture. Integration scaffolds and the full production stack remain works in progress.
03 / Engineering decisions
Make the boundaries visible.
Persistent tasks
Task state, sessions, execution events, and memory live in the database so operators can revisit the record after a run.
Approval boundaries
Policy evaluation, approval requests, and budget checks make the conditions around execution explicit.
Observable traces
The trace view records execution metadata, tool calls, summaries, and transitions that an operator can inspect.
Demo and live modes
Packaged demo data makes the dashboard explorable without every local service. Connected services supply live state when available.
04 / Current result
A working foundation, with a clear next step.
The repository includes the dashboard, runtime, security layer, MCP hub scaffolds, database schema, and Docker Compose configuration. Its README documents build, typecheck, test, and dashboard smoke-test commands.
Aegis is an alpha harness. Production work still includes broader authentication coverage, tenant isolation tests, richer approval lifecycles, secret storage, and backup and observability drills.