Your Agent Needs an SRE Handbook

Benchmarking extraction on 102 held-out contracts, the implementation compares full-document processing with RAG and examines MCP tool design, evidence validation, cost tradeoffs, and downstream agent accuracy.

Using executable checks on final database state as GRPO rewards, Microsoft’s ThinkingBox team lifted Qwen3.8-27B from 38 to 107 tasks passed on all 20 attempts and released the training code.

Using direct cache drop and reposition operations instead of history summarization, the proposed AKV system reports 27.7% lower costs and 1.7-point higher benchmark scores while improving serving performance.

Inspired by Anthropic’s Scion architecture, this open-source implementation combines sparse fetching, server-side history, batched change landing, and configurable conflict resolution to support multiple coding agents working on shared branches.

An AI research assistant recommended hiring another engineer; after consulting the Google SRE handbook, it instead called for an operational toil audit. That reversal frames a wider question: how should human expertise shape increasingly autonomous AI systems?

• Context engineering: Curated expert knowledge can redirect an LLM from surface symptoms to underlying organizational problems.
• Agent infrastructure: Inconsistent shell environments and tool configurations trigger repeated failures and wasted tokens; shared setups can preserve hard-won fixes.
• Open architectures: MCP, portable memory, and interoperable protocols offer ways for agents and people to cooperate across platforms.

Reliable AI depends on deliberately choosing the knowledge, constraints, and incentives guiding its work.

A computer-use benchmark reportedly achieved 18% accuracy on first attempts, reaching 33% only after retries. Those results frame a look at how agents select tools, navigate workflows, and verify generated code.

• Intent-based selection: Agents could describe desired actions and reasons to a classifier rather than access every tool directly, preserving an audit trail.
• Hybrid routing and WebMCP: Smaller LLM fallbacks can handle uncertain classifications, while websites expose different task-specific tools as workflows progress.
• Agentic verification: QA subagents can click through changed interfaces, record screenshots, and test outcomes before code reaches a pull request.

Reliable agent workflows need explicit intent, constrained actions, and evidence that intended outcomes occurred.

Getting stuck sooner helped keep engineers in control of an 800,000-line platform rebuild completed in six months. With Sonnet writing 60% of the project, the team prioritized human steering over longer autonomous coding runs.

• Model choice: Shorter reasoning loops surfaced questions earlier, especially around transaction engines and complex banking rules.
• ROI measurement: Dashboards compared token spend with shipped features and merged PRs, flagging discrepancies for discussion rather than judgment.
• Workflow design: A meta-harness moved Claude Code through refinement, specification, implementation, and verification, resetting context and skills between stages.

The practical lesson is to design agent workflows around visible decisions, measured outcomes, and controlled context.

Stopping a voice agent mid-sentence doesn't cancel a booking already sent to an external service. Production failures often begin when turn detection, playback, model generation, and application state fall out of sync.

• Latency tracing: Measure end-of-speech to first audio across endpointing, recognition, inference, tools, synthesis, and transport rather than blaming the model.
• Turn detection: Semantic and acoustic signals can distinguish thinking pauses from completed requests, balancing premature responses against unnecessary waiting.
• Interruptions: Truncate unplayed speech from history, distinguish backchannels from corrections, and track whether external operations remain active.

Reliable voice agents need playback-aware histories, explicit confirmation state, and traces that explain what actually happened.

Crash a coding agent halfway through a deployment plan, restart its worker, and pick up where it left off. An open-source goose–Dapr integration makes that possible by running agent work inside durable workflows rather than relying on saved chat history alone.

• Checkpointed turns: Dapr records completed conversation turns as workflow activities, avoiding replay and wasted model calls after recovery.
• Minimal integration: The diagrid-goose Rust crate registers goose with a Dapr Workflow worker and schedules multi-step tasks.
• Recovery boundaries: Persistent goose session storage is required; side effects within an interrupted turn still need idempotency protection.

Turn-level durability provides recoverable agent execution without rebuilding the existing tool stack.