leadbybuild.ingArchive

LeadByBuilding Newsletter · 8 min read

Enforce agent limits in the platform, expect red builds from agents, and check which clients still trust a revoked key

Two AI picks: the CNCF on bounding agents with identity, mesh and admission policy, and Thoughtworks on why a week of agent-only coding left builds red. Three engineering picks: Cloudflare's measurement of eMudhra key patching, Stripe's staged payouts rollout, and AWS's per-execution telemetry for durable Lambda functions.

AI & Agent Engineering

Agent securityverified

CNCF: bound agents with identity, mesh reach and admission policy, and deny them edit rights over their own limits

A CNCF post argues an agent's real authority is whatever its tools and credentials grant, and that this drifts without anyone deciding. Its Safe Operating Model uses rings: short-lived workload identity, a service mesh limiting reach, admission policy blocking unapproved tool bindings, and a gateway authorizing each call. The agent's identity must not be able to change those definitions.

Why it matters

Instructions state intent but tool bindings state capability, and capability is what executes. Treating an agent as a workload lets you reuse least-privilege controls you already run. This is an opinion post from two field CTOs, not a benchmark.

Do this week

List the tools and credentials each agent can use today and compare them with what you intended. Then check that no agent identity can write to its own tool allow-list or role bindings.

CNCF Blog →
Agentic developmentverified

Thoughtworks CTO: a week of agent-only coding reached 80% but kept the build red, so the last mile is the problem

Rachel Laycock describes a Thoughtworks experiment in which no human wrote code for a week and agents built an airline application. The team reached about 80% of the app, spent roughly two thirds of its time at the whiteboard and one third on the harness, and the build was frequently red. She cites DORA's 2025 research linking AI adoption to higher throughput but lower delivery stability.

Why it matters

If the build is red often, people stop trusting it and a green build proves less. Larger batches from agents also stress CI. This is one anecdotal experiment and the author's view.

Do this week

Track how often your main branch build is red and how long it stays red since agents were adopted. If it is rising, cap parallel agent work or fix the slowest check before adding more agents.

Martin Fowler (Rachel Laycock) →

Engineering

Securityvendor-reported

Cloudflare: forged certificates appeared within hours of the eMudhra key leak and many clients still trust the chain

Cloudflare logged about 22,000 handshakes using the weak eMudhra keys, two thirds to test sites; every handshake that verified certificates was refused. One forged certificate was dated October 20 at 18:19 UTC, under an hour after Mozilla's bug report. In 50 million challenge-page probes, Android WebView was 30 to 40% still accepting, and over half of probes on large networks accepted the chain.

Why it matters

Revocation by itself does not protect users who do not update, and bots and embedded browsers lag. The probe data is skewed by bot traffic, as Cloudflare says. The AI-assisted source review covering all systems took four hours.

Do this week

Check that your outbound clients, base images and embedded WebViews have the eMudhra distrust update, and add a test that your trust store can be changed without a full release.

Cloudflare Blog →
Migrationvendor-reported

Stripe moved $1B of vendor payments in four waves and found bad recipient data caused most failures

Stripe moved its own vendor payments to Global Payouts: one low-value payment per US flow in February 2026, then four waves from March, raising amounts only after nearly all payouts in the prior wave succeeded. It reports over $1 billion paid to 3,400 vendors, market setup cut from six weeks to one, and 99.9% success. Bad recipient data caused most failures, and ACH and wire limits rose to $100 million.

Why it matters

The method is reusable: ramp by value with a success gate, move validation before the irreversible step, and test size as well as volume. These are Stripe's own figures about its own product.

Do this week

For your next migration of a money-moving or other risky flow, define the success threshold that opens each wave. List the top failure causes and move those checks ahead of the call that cannot be undone.

Stripe Blog →
Observabilityverified

AWS: Durable Execution Telemetry Exporter emits one queryable record per Lambda durable execution

AWS released an open source plugin for Lambda durable functions that builds a cumulative JSON record per execution, with status, timing, input, output and each step. The default writes to the function's own CloudWatch log group; 13 exporters cover S3, DynamoDB, EventBridge, OTLP and others. emitMode defaults to on-complete. AWS also released an MCP server so a coding agent can query the data.

Why it matters

A durable execution spans many invocations and log streams, so per-invocation logs cannot say which step fails most. One record per execution turns that into a query. Input and output fields may carry sensitive data, so review what you emit.

Do this week

If you run durable functions, enable the default exporter on one workflow and run a Logs Insights query for the most-failed step. Decide which input and output fields are safe to emit first.

AWS Compute Blog →

Worth experimenting with

Audit what one agent can do versus what you meant it to do

Pick one agent in use. List every tool, credential and delegation it holds, mark each as intended or accumulated, and remove the accumulated ones. Then confirm the agent's own identity cannot edit that list.

Why. The CNCF post says agent reach drifts through changes nobody decided on. A one-off audit shows how large the gap is for you.

Measure. Count of tools and credentials held versus intended before and after, and whether the agent identity can write to its own permission objects (yes or no).

Classic / essential read

Microservices: a definition of this new architectural term

James Lewis and Martin Fowler · 35 min

Written in 2014, it lists common traits of the style: componentization via services, smart endpoints and dumb pipes, decentralized data, infrastructure automation and design for failure.

One lesson. Weigh each trait as a cost you take on, such as failure handling and automation, before splitting a service, rather than adopting the label.

Read the original →

Put limits where capability lives, not in instructions. Measure how fast fixes reach clients, and keep your build trustworthy as agents raise the volume of change.

Get this every morning.

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

Subscribe