leadbybuild.ingArchive

LeadByBuilding Newsletter · 8 min read

Measure delivery flow, not AI usage, and write the rules that keep agent history and checklists honest

Three AI picks: LeadDev on Atlassian's survey of AI gains lost to code review, Spring AI keeping reasoning data intact across tool calls, and a Claude skill that runs the AWS Agentic AI Lens. Two engineering picks: ECS Managed Instances repacking tasks to raise utilization, and a telemetry exporter for Lambda durable functions.

AI & Agent Engineering

Engineering leadershipverified

LeadDev: in Atlassian's survey, developers saving 10 hours a week with AI fell from 68% to 25%

LeadDev reports on Atlassian's 2026 State of Developer Experience survey of 1,155 developers and 405 engineering leaders. 87% of developers say AI helps them ship code faster, yet 72% report the same or more friction across the lifecycle. 65% regularly spend time reviewing, debugging and correcting AI-generated code. Only 17% of developers say their organization has a clearly communicated and implemented AI strategy.

Why it matters

The queue moves to the next stage: review, merge conflicts and tests. Per the report, 60% of developers say AI usage is treated as a productivity signal, while only 42% of leaders think they have the right metric. These are survey answers, not measured delivery data.

Do this week

Stop counting prompts or accepted suggestions as productivity. Pick one flow metric, such as PR review wait time, and track it for four weeks alongside AI usage on one team.

LeadDev →
Agent frameworksverified

Spring AI 2.1 RC models messages as ordered parts so reasoning signatures survive tool-calling round trips

Spring AI 2.1.0-RC1 changes the message data model: each message holds an ordered list of parts, such as text, reasoning and tool calls, including opaque payloads like Anthropic signatures and OpenAI encrypted reasoning items. The post says dropping, reordering or editing them can cause an HTTP 400 or a model that quietly loses earlier reasoning. Its rules: keep the order, replay verbatim and persist the parts, not the text.

Why it matters

If you store chat history as plain text, reasoning models may lose context across tool rounds. The post notes histories stored before 2.1 have no reasoning parts, ChatMemory is deprecated, and signed reasoning is not sent to another provider.

Do this week

Check how your agent code stores history for reasoning models. If it keeps only text, test a multi-step tool call and a restart, and keep provider payloads unchanged.

Spring Blog →
Agent engineeringvendor-reported

Smarsh turned the AWS Agentic AI Lens into a Claude skill that reviews 150 practices against a repo

AWS describes how Smarsh made the 150-practice Agentic AI Lens executable as a Claude skill. It holds the lens as JSON, a rubric where every status must cite repository evidence or be marked Not Reviewed, and a script that rejects unknown IDs. On two synthetic repositories, runs with the skill passed 16 of 16 expectations and runs without it 8 of 16. Smarsh says reviews take hours, not days.

Why it matters

The useful part is the guardrails: evidence per finding, a fixed source of truth and a script that fails on invented items. The test set is small and synthetic, and the post is written with a partner, so treat the results as vendor-reported.

Do this week

Pick one checklist your team reviews by hand. Put it in a JSON file, require cited evidence for every answer and fail the run on unknown IDs, then compare with a manual review.

AWS Architecture Blog →

Engineering

Cloud costvendor-reported

ECS Managed Instances now repacks tasks across several underused instances and consolidates Spot capacity

Amazon ECS Managed Instances adds cross-instance consolidation, which evaluates tasks across several underutilized instances together instead of one at a time, and Spot consolidation, which checks replacement capacity first. In AWS's test with 261 tasks, the fleet fell from 19 instances to 13, CPU utilization rose from 42% to 91% and memory from 41% to 88%. It is free and applies to existing capacity providers.

Why it matters

Narrow instance-type ranges gave the old per-instance logic few chances to save money. Because it applies automatically to existing providers, task movement may increase. The figures are one AWS test, not a general result.

Do this week

Check your capacity providers for low CPU and memory utilization. Confirm your tasks handle graceful replacement and have health checks and drain settings before relying on consolidation.

AWS What's New →
Observabilityvendor-reported

Lambda durable functions get a telemetry exporter with redaction, sampling and 13 built-in destinations

AWS released the Durable Execution Telemetry Exporter plugin for Lambda durable functions. It captures execution status, operation timing, step input and output, and errors, and sends them to 13 built-in destinations such as CloudWatch Logs, S3 for Athena and DynamoDB. You can redact sensitive fields and set a sampling rate. It supports TypeScript, Python, Java and Go and costs nothing extra. A companion MCP server lets agents query the data.

Why it matters

Durable executions can run up to one year, longer than standard retention, so exporting history matters for audits and debugging. Step input and output can contain personal data, so redaction rules are a design decision, not an option.

Do this week

Before enabling the exporter, list which step inputs hold sensitive data and write the redaction rules first. Start with a low sampling rate and compare storage cost with the debugging value.

AWS What's New →

Worth experimenting with

Measure review wait time on one team for four weeks

Pick one team. Record the time from PR opened to first review and to merge for every PR over four weeks, and note which PRs were mostly agent-written. Compare with the four weeks before if you have the data.

Why. LeadDev's piece suggests AI time savings move into review. Your own numbers show whether that is happening on your team before you change process.

Measure. Median hours to first review and to merge, plus the share of PRs over 500 changed lines, before and after. If neither moves, the survey result may not describe your team.

Classic / essential read

Blue Green Deployment

Martin Fowler · 4 min

Fowler describes keeping two near-identical production environments, testing in the idle one and switching the router, with a fast rollback. He notes that databases are the hard part and advises separating schema changes from application upgrades.

One lesson. Whenever you plan a risky change, ask two questions: what is the one-step rollback, and what happens to writes that land during the switch? Fowler says the database is where this gets hard.

Read the original →

Commute watch · optional

The Production AI Playbook: Deploying Agents at Enterprise Scale — Sandipan Bhaumik, DatabricksAI Engineer · 37:06Jack Vanlightly — A systematic approach to building reliable distributed systemsHeisenbug — конференция по тестированию · 59:02

Count what ships, not what is generated: track review wait time, keep reasoning data and checklist evidence intact, and know how a consolidation or cutover rolls back.

Get this every morning.

Free. Same issue for everyone. One click to unsubscribe.

Subscribe