AgentBag vs. an agent framework
Frameworks solve orchestration: how a model plans, calls tools and loops. That is a real problem and they solve it well. It is also not the problem that stops agents reaching production, which is almost always mail, money, permissions and audit.
What it is for
an agent framework:How the agent thinks: planning, tool loops, retries, evaluation.
AgentBag:What the agent has: mailbox, ledger, corpus, permissions, audit trail.
Identity
an agent framework:The agent is a process. It borrows your credentials.
AgentBag:The agent is an addressable principal with its own mailbox, its own scope and its own revocation.
Audit
an agent framework:Traces for debugging a run.
AgentBag:A durable Postgres task ledger with delegation edges, artifacts and recorded interventions.
Tool protocol
an agent framework:Usually MCP client, sometimes.
AgentBag:MCP on both ends: client and OAuth 2.1 server.
Our MCP client accepts a static bearer token for custom remote servers; full OAuth 2.1 there is planned, not shipped.
Maturity of the reasoning layer
an agent framework:Years of iteration, large communities, deep eval tooling.
AgentBag:Not our layer. Bring your own.
If you are choosing on the strength of the orchestration loop, pick the framework.
Rows marked in the third column with a note are the ones where we are honestly not better. Every comparison on this site has at least one (see the full capability ledger for the rest).