The infrastructure agents need to act on real systems — identity, runtime, safe execution, memory, skills, and open protocols — already shipped and merged in production.
Every claim links to a merged pull request — github.com/ahmedhesham6.
Six layers. Each one thing agents need in production — and the thing I shipped for it.
An agent proves who it is on every call — and never holds the master key.
Credentials are issued at runtime, expire within the hour, and every action is audited with secrets redacted. A production implementation of the open Agent Auth Protocol; observability tools were only the first connectors.
The protocol and its APIs landed in a single +14k-line pull request.
An agent can run for weeks, pause for a human, and resume exactly where it stopped.
Durable sessions with checkpoints and replayable event logs, reachable from Slack, Discord, or Telegram through one gateway — with human approvals in the loop.
Agents do real work inside sealed sandboxes — and secrets never leave with them.
Work runs in sandboxed containers with the agent's knowledge mounted in; secrets are redacted span-aware before anything leaves; subagents are routed to the right model deliberately, not by guesswork.
Agent memory that stays fast when the load is real.
Concurrency fixes merged into a 26k-star context database, plus a persistent knowledge store agents can search, read, and write — mounted straight into their sandboxes.
Agent skills that install like packages — publish once, install anywhere.
Global or project scope, across Claude Code, Codex, Gemini CLI, OpenCode, and Amp. Founding contributor; authored PR #1.
Any editor, any agent — the same session semantics.
Authored the session/list RFD and its implementation, plus the usage and context-status RFD, in the cross-editor Agent Client Protocol (3.5k stars).
Ahmed has already shipped the production agent-infrastructure layer: verifiable agent identity, scoped capability execution without stored secrets, durable multi-channel sessions, sandboxed knowledge, distributable skills, and open agent protocols — AAP for agent auth, ACP for editor-agent sessions.
Every claim traces to a merged PR under github.com/ahmedhesham6 · verified July 2, 2026