[bmdpat]
All writing
4 min read

agent-sre on PyPI: what SRE for AI agents actually means

Microsoft just shipped agent-sre on PyPI. Seven packages: SLOs, error budgets, circuit breakers. Here is what it does, what it does not, and why solo builders still need agentguard47.

#agent-sre#agentguard#sre#ai-agents#python
Share LinkedIn

TL;DR

  • agent-sre landed on PyPI as part of Microsoft's Agent Governance Toolkit. Seven packages. SLOs, error budgets, circuit breakers, chaos testing, progressive delivery.
  • The circuit breaker trips when an agent's safety SLI drops below 99%. This is designed for teams running dozens of agents at scale, not one person with Task Scheduler and a markdown vault.
  • agentguard47 solves a smaller problem. Wrap a function with @guard(budget_usd=0.10) and it throws at the function boundary if spend hits the limit. No SRE background. No config file. No service to run.
  • Most solo builders are not running agent fleets. They run one agent that calls Claude or GPT in a loop. The risk is a runaway loop that charges $80 while they are asleep, not a 99% SLI miss.

Cap agent run cost with AgentGuard

agent-sre just landed on PyPI as part of Microsoft's Agent Governance Toolkit. Seven packages. SLOs, error budgets, circuit breakers, chaos testing, progressive delivery.

That is the full SRE playbook ported to agent systems. It is a real idea and it deserves a real look.

I want to talk about what it actually means for solo builders, because the approach is meaningfully different from what I built with agentguard47.

What agent-sre does

Microsoft's toolkit applies org-scale SRE to agent fleets. The circuit breaker trips when an agent's safety SLI drops below 99%. The error budget engine tracks burn rate across an entire deployment. Chaos testing stress-tests failure modes before production.

This is designed for teams running dozens of agents at scale. Think: enterprise ML platform team with dedicated SRE headcount, not one person with a Task Scheduler and a markdown vault.

To use it well you need a defined agent fleet, SLI instrumentation, a policy engine, and someone who speaks SRE. That is a real engineering investment. The tooling is sophisticated because the problem it targets is sophisticated.

What I built instead

agentguard47 solves a smaller, more immediate problem.

I was burning money because a single agent function had no budget ceiling. No fleet. No policy engine. Just: I need this function to stop if it hits $0.10.

@guard(budget_usd=0.10)
async def research_competitors():
    ...

That is the whole API. One decorator. Framework-agnostic. Throws at the function boundary if spend hits the limit. No SRE background required. No config file. No service to run.

The Cost Guard component inside agent-sre works at the org level. AgentGuard works at the per-function level. These are not competing solutions. They solve at different layers.

When to use which

agent-sre is the right tool if you are running a multi-agent fleet with policy requirements, have a team that already speaks SRE, and need chaos testing and staged rollouts.

agentguard47 is the right tool if you are one person with one agent and one credit card, you want enforcement in one decorator with no config, or you are prototyping and need a hard stop before you accidentally charge $200 in a test run.

The honest version: most solo builders are not running agent fleets. They are running one agent that calls Claude or GPT in a loop. The operational risk is not a 99% SLI miss. It is a runaway loop that charges $80 while they are asleep.

agentguard47 is a pip install away from fixing that specific problem.

The category is real

More tooling in this space is a good sign. The fact that Microsoft shipped seven packages targeting agent observability and safety validates the problem space. Costs run away. Agents behave unexpectedly. Runtime enforcement matters.

Solo builders just need a different entry point than enterprise SRE tooling.

If you are past the "oops I spent $50 on a test" phase and running a real fleet, go look at agent-sre. The Microsoft toolkit is open source and legitimately well-designed.

If you are still in the "I do not want to get surprised by my bill" phase, agentguard47 is one install.

pip install agentguard47

Full docs: https://bmdpat.com/tools/agentguard

FAQ

What does Microsoft agent-sre actually do?

Org-scale SRE for agent fleets: SLOs, error budgets, circuit breakers, chaos testing, and progressive delivery across seven packages.

When is agentguard47 the better first tool?

One person, one agent, one credit card. You want enforcement in one decorator, or a hard stop before a test run charges $200.

Are they competing products?

No. Cost Guard in agent-sre works at the org level. AgentGuard works at the per-function level. Different layers.

Get the Local AI Field Kit

Four copy-ready tools now, then one evidence-backed Local AI Lab Note on Friday when there is something worth sharing.

Try the free agent run check first

Get the requested artifact now, then at most one evidence-backed Local AI Lab Note on Friday when there is something worth sharing. One-click unsubscribe. No sponsored placements. Privacy.

PH

Patrick Hughes

I build BMD and publish measured AI runs, failure reports, and reusable checks. Nashville, Tennessee.

More writing