Writing about MCP, agents, and the execution layer.

Engineering deep dives from the Harbor team on how teams of AI agents share tools, credentials, and audit. New posts ship as ideas land.

Harbor Internals, Part 1: why exec does not mean secret access

Harbor exec hands code capabilities, not credentials. A walk through the secret boundary: isolate split, secret refs, mediated outbound fetch, error sanitization, and sealed CLI secrets.

Author: Umang Bhalla
Published: May 30, 2026
Estimated reading time: 11 min read

security
execution layer

Read article

More writing

04 posts

May 26, 2026

What is an execution layer?

By Umang Bhalla

An execution layer turns submitted code plus workspace context into an audited run: source resolution, MCP pooling, sandbox isolate, orbit primitives.

Read article

May 26, 2026

Shared workspaces for AI agents

By Harbor team

Why N agents × M tools needs a shared workspace, not per-agent toolboxes — credentials, memory, runs, and audit live with the team.

Read article

May 26, 2026

Control plane vs MCP gateway: when each one fits

By Harbor team

MCP gateways forward tool calls. A control plane owns the workspace around them — discovery, credentials, execution, audit. When each one fits.

Read article

May 15, 2026

Code Mode and why shell exec is on borrowed time

By Nishank Jain

Code Mode moves agent execution from turn-by-turn tool calls into typed programs that can run inside an isolated runtime. The article frames this as the next execution layer for agents: fewer model round trips, less context waste, and more composable workflows across MCP tools, CLIs, APIs, and GraphQL.

Read on X