Apache Burr: Build reliable AI agents and applications

Original Article ↗ Hacker News Discussion ↗

Apache Burr, a new Apache-incubated framework for building AI agents and state-machine-based workflows, is drawing attention both for its technical goals and its highly stylized, AI-generated-looking website. Commenters compare it to other agent stacks like LangGraph, Strands, Pi and bespoke “hand-rolled” agents, debating whether general-purpose frameworks help or just obscure core logic, and arguing that the real value lies in observability, orchestration, and monitoring rather than the agent loop itself. Many see Burr’s unopinionated, lower-level approach and built-in tracing as promising, but others question whether any agent framework meaningfully improves reliability given current model limitations and the crowded ecosystem.

Project Purpose, Origin & Naming

  • Burr is an Apache-incubated framework for orchestrating AI agents via state machines.
  • It grew out of earlier work on Hamilton (a DAG-based library) as a way to manage state between DAG executions, especially where cycles/recursion are needed.
  • It is intentionally low-level and “bring your own functions/classes,” aiming to be unopinionated orchestration rather than a full agent stack.

Positioning vs Other Frameworks

  • Compared to dspy.ai: described as more low-level, not a direct competitor.
  • Compared to LangGraph: some see Burr as very similar, “LangGraph with a builder pattern.”
  • Compared to Strands / AgentCore: those are perceived as more opinionated and tied to specific clouds; Burr is evaluated as a potential alternative if it matures.
  • Other tools mentioned in the same space: Pi, NanoBot, Nvidia Openshell, Codex, OpenClaw, Jido, Forge; Burr is seen as one more entry in a crowded market.

Frameworks vs Hand-Rolled Agents

  • Strong current that simple agents are easy to write directly (loop + tools + context + parsing), and bespoke code is often clearer and more maintainable.
  • Counterargument: reinventing primitives like tool schemas, serialization, and harness logic is wasteful; using a harness is akin to using an API client instead of reimplementing a protocol.
  • Consensus that the real value of frameworks is not the basic agent loop, but:
    • Observability/tracing
    • Guardrails and policy
    • Monitoring, deployment, versioning, evals, A/B testing.

Reliability, Orchestration & Context

  • Some argue “reliable” agents are mainly about decomposition, orchestration, and context management, not just state machines.
  • Others are skeptical there is such a thing as fully “reliable AI,” or that Burr’s approach meaningfully solves that.
  • Several posts emphasize:
    • Multi-step workflows with classifiers, tools, approvals, and recursion quickly become complex.
    • Context management, long-term memory, and “brains” (e.g., file-purpose summaries, blast-radius analysis) are key to scalable agents.
    • Agent swarms and spec-driven development are proposed as separate reliability strategies.

UI, Community & Perception

  • Many criticize the landing page as “vibe-coded” / performative UI: gradients, animated buttons, Tailwind-style template, JavaScript-heavy.
  • Some feel this aesthetic and a Discord-centric community make the project look rushed or unserious for Apache.
  • Others note the site was user-contributed and not representative of core technical quality.

Open Questions Raised

  • How Burr handles agent authentication and protocols like MCP is unclear from the docs.
  • Questions about comparisons to Pydantic and detailed security/auth patterns go unanswered in the thread.