Building Agent-im: A Four-Layer Model for Agent Versioning, Memory and Secrets
yangyi · x · 2026-11-17
The team behind NewMax Agent-im lays out a systematic take on agent version management, arguing an "Agent GitHub" will emerge but look more like a Docker Registry + branchable database + enterprise identity system + CI hybrid.
Key points:
- A working agent is more than a prompt: system config (soul.md), skill/tool dependencies, access boundaries, secrets, and evolving memory (shared/specialized/session) all need distinct handling
- The core scheme splits agents into four layers: definitions (Git-style diffing), dependencies (package-manager-style link + lockfile), state (data versioning with provenance), environment (declared needs, never versioned)
- Golden rule: secrets and permissions must never be forked — agents only declare needed credentials, the deployer's environment binds them
- Docker-style layered stacking (base → specialty → team → instance) defines the link vs fork boundary
The piece also derives seven infra layers for enterprise "Agent Resources" and argues this should become an open protocol.
More from AGI Musings
- Mark Cuban: slow enterprise AI adoption is why the "trillionaire in the basement" is coming — alvelda · 2026-11-17
- Sam Altman imagines AI compressing 10–100 years of scientific progress into a single year — kimmonismus · 2026-11-17
- Dan Shipper's 'proof of work': rank feeds by how many tokens agents spend on content — danshipper · 2026-11-17
- Success or failure, Anthropic spells crisis — unless Jevons paradox saves the economy — sebpaquet · 2026-11-17
- Biggest U.S. data centers emit 25.7M tons of CO2 a year, rivaling entire nations — asusarla · 2026-11-17
- Perplexity CEO: AI for automating scientific research is the best compute investment — dr_alphalyrae · 2026-11-17