XNews

AI agents and context-graph infrastructure

AI agents and context-graph infrastructure

Two threads detail agent architectures and infrastructure. - MiroThinker 1.5 presented as open-source research agent - Model uses 30B parameters yet outperforms much larger models - Inference cost cited at $0.07 per run - MiroThinker automates search, hypothesis, evidence gathering, verification - Supports up to 400 tool interactions per task in v1.5 - Memory-managed multi-round reasoning retains recent 5 contexts - Glean CEO Arvind Jain argued 'context graphs' are essential for enterprise agents - Context graph maps people, actions, and relationships across tools - Glean focuses on capturing 'How' (operational traces) rather than 'Why' - Technical challenges: full observability, atomic activity data, signal extraction from noise - Required stack: connectors, indexes, graphs, memory for agent execution

Key insights

MiroThinker 1.5 positioned as an efficient open-source research agent

MiroThinker 1.5 is presented as a high-performing open-source research agent using a 30B-parameter model that reportedly outperforms models at the 1T-parameter class. The project emphasizes low inference cost (cited at $0.07), automated workflows that search, hypothesize, gather evidence, verify, and iterate, multi-step self-correction, long-context memory, and the ability in v1.5 to call up to 400 tools per task to support complex, hierarchical reasoning chains.

MiroThinker implements memory-managed multi-round reasoning and tool orchestration

MiroThinker uses memory-managed multi-round reasoning, retaining the five most recent contexts while cleaning irrelevant history to maintain information integrity and low resource usage. The agent supports automated information discovery, multi-step planning with self-correction, and tool invocation/execution across extensive sequences, enabling recursive and hierarchical reasoning for research-grade tasks.

Context graphs framed as the next-generation data platform for enterprise agents

Glean CEO Arvind Jain argued that enterprise agents require 'context graphs'—maps of people, activities, and relationships across systems—to understand real work processes. He outlined that context graphs capture operational 'How' traces left by edits, approvals, and communications, enabling agents to infer business intent from process signals and navigate complex workflows.

Building context graphs requires deep observability and atomic activity data

Arvind Jain identified three technical barriers to constructing context graphs: full-chain observability across applications, recording time-stamped atomic activity data (edits, comments, messages), and extracting signal from noisy fragments to recognize cohesive work events like customer onboarding or product launches. He outlined a stack of connectors, indexes, graphs, and memory to support agents executing tasks reliably.

High tool-invocation capacity enables recursive research workflows

MiroThinker v1.5's capability to perform up to 400 tool interactions per task is positioned as a mechanism for building recursive, layered reasoning chains: plan → search → verify → revise → repeat. This capacity supports sustained autonomous inquiry across diverse information sources and toolsets.

Memory policies emphasize recency and signal preservation

MiroThinker applies memory management that preserves the most recent five conversational contexts while pruning irrelevant history to keep resource usage low and maintain information integrity, supporting iterative verification without retaining excessive noise.

Source posts