AgentsEvery agents story.
4 stories across the archive, newest first.
Group Relative Policy Optimization samples several answers per prompt, scores them, and pushes the model towards the ones that beat the group average. No value network is needed, which saves a model's worth of memory and made large-scale RL on reasoning practical for DeepSeek.
Why you ·Your agent work at HivePro and Susanoo lives on tool-use and reasoning quality. GRPO is the cheapest credible way to RL-tune a small model on a verifiable task, such as a JanusGuard triage policy scored against labelled CIC-IDS2017 flows.
Open the deep dive: GRPO, explained: reinforcement learning without a critic →A durable execution engine persists the result of every side-effecting step as an event. If the worker crashes, another worker replays the history, skips the steps already done, and carries on from where it stopped. For agents that call slow, paid LLM APIs and tools, that turns crashes from expensive into boring.
Why you ·You already run Temporal next to LangGraph at HivePro. Knowing exactly what replay does (and why workflow code must be deterministic) is the difference between a resilient agent platform and mysterious non-determinism errors in production.
Open the deep dive: Durable execution: why long-running agents need a workflow engine →When an agent pulls in a web page, an email or a tool result, any text inside it lands in the same context window as your instructions. An attacker who controls that content can try to steer the agent. The defences are architectural: least privilege, separating data from instructions, and human confirmation for risky actions.
Why you ·Every tool you add to the HivePro agent platform or Susanoo is a new input channel an attacker might control. JanusGuard also needs this threat model if it ever reads attacker-supplied logs or payloads.
Open the deep dive: Prompt injection through tool outputs →Vector search finds passages that mean the same thing; keyword search finds exact names, error codes and IDs. Postgres can do both: built-in full-text search and the pgvector extension. Run both queries and merge the ranked lists, for example with reciprocal rank fusion, to get memory that handles both kinds of question.
Why you ·Hybrid RAG is already part of your HivePro stack, and doing it inside Postgres builds directly on the Postgres internals you're learning. It's also a cheap memory layer for Susanoo before you need a dedicated vector database.
Open the deep dive: Postgres + pgvector hybrid search as agent memory →