Isolation primitives, memory graphs, and durable control planes for agent loops

polos-dev

Open-source agent runtime with Docker or E2B sandboxes, durable workflows that resume mid-step, HITL approvals, Slack triggers, and OpenTelemetry tracing.

Use caseUse it to ship production agents that survive crashes and keep costs low via prompt caching.

Graphicle-AI

Single Rust binary that builds a tree-sitter code graph plus project memory (docs, decisions, sessions) and serves cited excerpts over MCP within a token budget.

Use caseUse it to give any MCP coding agent precise file-and-line context instead of dumping whole files.

rivet-dev

Durable actor runtime plus sandbox-agent server so coding agents run inside isolated boxes while the harness and memory live outside, controlled over HTTP/SSE.

Use caseUse it to keep the agent loop durable and auditable while tools execute in disposable sandboxes.

liliang-cn

Rust framework that turns models into production agents with typed ReAct loops, staged autonomy levels, guards, and built-in observability.

Use caseUse it to build domain-specific agents with compile-time safety and progressive trust gates.

graph-memory

MCP server that indexes docs, code, and files into a semantic graph and exposes dozens of tools for search, knowledge, tasks, and skills.

Use caseUse it to give coding agents a shared project knowledge graph with REST and web UI.

memory-graph

Graph-DB MCP memory server for coding agents with relationship tracking, typed stores, and session briefing commands.

Use caseUse it to persist patterns, fixes, and decisions across agent sessions with explicit links.

ag-harness

Rust ADE utility for structured LLM turns with deny-by-default tools, durable sessions, and native sandboxed bash.

Use caseUse it to enforce typed, grant-gated agent turns with Seatbelt or Landlock isolation.

Sahajmeet Kaur / TrueFoundry

The post ranks open harnesses and highlights TrueForge's sandbox-as-tool design that cut per-run cost versus managed agents on the same tasks. It argues the harness layer now matters more than the base model for production reliability.

Read