Busabase x Knowledge Base for Deep Agents CodeKnowledge Base for Deep Agents Code
Deep Agents is an agent harness for long-running work. It combines a virtual filesystem, memory, skills, context offloading, subagents, and human approval. Those layers make work durable, but each layer has a different scope and permission boundary.
MAP
Long-running context is useful only when its storage, permissions, and handoff remain inspectable.
Deep Agents is an agent harness for long-running work. It combines a virtual filesystem, memory, skills, context offloading, subagents, and human approval. Those layers make work durable, but each layer has a different scope and permission boundary.
Read, write, edit, glob, grep, execute
Ordered read/write allow and deny rules
Always-loaded memory and on-demand procedures
Fresh isolated context and one final report
BOUNDARY
Memory is always loaded while Skills use progressive disclosure
Long-running context is useful only when its storage, permissions, and handoff remain inspectable.
| Virtual filesystem | Read, write, edit, glob, grep, execute | Context and working artifacts |
|---|---|---|
| Filesystem permissions | Ordered read/write allow and deny rules | The agent’s access boundary |
| Memory and Skills | Always-loaded memory and on-demand procedures | Different context lifecycles |
| Subagent handoff | Fresh isolated context and one final report | Delegation evidence for the parent |
SCOPE
Filesystem permissions use ordered first-match rules
Deep Agents is an agent harness for long-running work. It combines a virtual filesystem, memory, skills, context offloading, subagents, and human approval. Those layers make work durable, but each layer has a different scope and permission boundary.
Context and working artifacts
The agent’s access boundary
Different context lifecycles
Delegation evidence for the parent
EVIDENCE
Subagents return a final report from isolated context and can be restricted independently
Memory is always loaded while Skills use progressive disclosure Filesystem permissions use ordered first-match rules Subagents return a final report from isolated context and can be restricted independently
| Virtual filesystem | Read, write, edit, glob, grep, execute | Context and working artifacts |
|---|---|---|
| Filesystem permissions | Ordered read/write allow and deny rules | The agent’s access boundary |
| Memory and Skills | Always-loaded memory and on-demand procedures | Different context lifecycles |
| Subagent handoff | Fresh isolated context and one final report | Delegation evidence for the parent |
ACCEPTANCE
Long-running context is useful only when its storage, permissions, and handoff remain inspectable.
Memory is always loaded while Skills use progressive disclosure Filesystem permissions use ordered first-match rules Subagents return a final report from isolated context and can be restricted independently
Verify Read, write, edit, glob, grep, execute before accepting Context and working artifacts.
Verify Ordered read/write allow and deny rules before accepting The agent’s access boundary.
Verify Always-loaded memory and on-demand procedures before accepting Different context lifecycles.
Verify Fresh isolated context and one final report before accepting Delegation evidence for the parent.
FAQ
Knowledge Base for Deep Agents Code: practical questions
Long-running context is useful only when its storage, permissions, and handoff remain inspectable.
Does memory have the same lifecycle as a Skill?
No. Memory is loaded at startup; Skills are loaded progressively when needed. Their freshness and review paths differ.
Do filesystem permissions validate content?
No. They restrict read and write operations. A permitted file can still contain stale or incorrect claims.
Is a subagent report automatically accepted?
No. It is a handoff from isolated context. The parent or reviewer must verify the result.
Why preserve the virtual filesystem boundary?
It makes intermediate artifacts, permissions, and context recovery inspectable across long tasks.
NEXT
Long-running context is useful only when its storage, permissions, and handoff remain inspectable.
Deep Agents is an agent harness for long-running work. It combines a virtual filesystem, memory, skills, context offloading, subagents, and human approval. Those layers make work durable, but each layer has a different scope and permission boundary.