The differentiator is no longer the model — it is the harness: 39 documented initiatives show five companies turning agent access into an engineered software factory

Asked:
Detailed Report of Agentic Engineering and Harness engineering at Zalando , Uber , Netflix, AirBnB, Anthropic and others

This report covers 51 publicly documented agentic and harness engineering initiatives current to 2 September 2026: 39 at the five focal companies (Zalando, Uber, Netflix, Airbnb, Anthropic) and 12 more from Block, Google, GitHub, Stripe, Microsoft, LangChain, METR and research groups — each with lifecycle scope, harness detail, metric where one was published, and source tier. Counts are documented public examples, not totals of internal systems, and are not comparable across companies.

Company × lifecycle coverage: who has published what, and where the metrics are

documented initiatives (darker = more)green figure = initiatives with a public metrichover a cell for the initiatives and headline numbers behind itempty = nothing published in that slice, not zero activity

Definitions, precisely

Agentic engineering is the organizational and software-delivery discipline of structuring work so agents can plan, act with tools, inspect feedback, and iterate toward an outcome. Harness engineering is the surrounding execution and control layer that turns model capability into dependable work: environment, tool interface, context/memory, orchestration, observability, verification/evaluation, and governance/security. Not every initiative below is fully autonomous — the record spans pipeline automation (Zalando's migration tool), interactive copilots (Gemini CLI, Claude Code), background agents (GitHub's coding agent, Headless Goose), and managed software factories (Uber, Stripe Minions). They are included because they show the evolutionary path toward agentic systems.

Where the documented examples cluster across the lifecycle

documented exampleswith public metricscounts of documented public examples across all 11 organizations — not total internal systems
Platforms, context and evaluation 11 companies
16 documented · 6 with metrics · medium.com
Coding and development 7 companies
16 documented · 3 with metrics · markolukicic.substack.com
Test, validation and optimization 6 companies
10 documented · 7 with metrics · medium.com
Review and governance 6 companies
8 documented · 4 with metrics · careers.airbnb.com
Operations and incident response 2 companies
5 documented · 2 with metrics · goo.gle
Migration and refactoring 1 companies
1 documented · 1 with metrics · engineering.zalando.com

What the matrix cannot say alone

Testing and validation is where metrics are densest — 7 of 10 documented initiatives carry a number — because tests give agents deterministic ground truth: DragonCrawl blocked ten high-priority bugs in three months, and PerfInsights' dual validation cut false positives from over 80% to the low teens. uber.com
The cheapest headline win in the dataset is Zalando's UI library migration — about 90% accuracy at under $40 per repository — while its risk bot now auto-approves 33% of PRs and its LiteLLM proxy serves 2k monthly active users on six small pods. engineering.zalando.com
Airbnb's strongest primary evidence is a single migration: nearly 3.5K Enzyme test files converted in six weeks versus an estimated 1.5 engineer-years by hand, with one stage moving 75% of target files in four hours through per-file parallelism, retries and feedback loops. medium.com
Anthropic's only public metric is about governance, not throughput — OS-level sandboxing cut permission prompts 84% — and Netflix's oci-agent estimate reaching just 25% of a polished baseline is the dataset's clearest warning that generalist agents need specialist paved paths. anthropic.com
Uber's biggest 2026 numbers — 5,000+ tests/month, ~1,800 changes/week at 95% engineer reach, 7x user and 9.4x request growth — are all secondary-source claims; its verified primary figures sit in testing and optimization. newsletter.pragmaticengineer.com

Company deep dives

Zalando — Layered platform, PR-risk automation and SRE pipelines

9 documented initiatives · 5 with public metrics · 7 lifecycle scopes

Zalando's public record shows a layered internal platform: a LiteLLM API proxy serving 2k monthly active users on just six small pods, chat and CLI surfaces, and production pipelines on top. A risk-based bot auto-approves 33% of PRs classed as low-risk; an LLM-powered UI component library migration hit about 90% accuracy at under $40 per repository; and the SRE postmortem pipeline delivered a 3x productivity boost in summaries and processed annual analysis in under 24 hours. An LLM-as-a-judge framework assures search quality across languages before market launches.

Uber — Broadest measured factory-style portfolio — with secondary-source caveats

7 documented initiatives · 7 with public metrics · 7 lifecycle scopes

Uber's primary engineering posts document DragonCrawl blocking ten high-priority bugs in three months while saving thousands of developer hours, and PerfInsights' dual validation (with LLMCheck) cutting false positives from over 80% to the low teens, plus a 33.5% drop in validated static-analysis findings in four months. Newer scale figures — 5,000+ unit tests/month from Autocover, about 1,800 code changes a week and 95% engineer reach for Minion, 7x weekly active user and 9.4x request growth for the Software Factory — come from secondary reporting and should be treated as claims, not verified primary figures.

Netflix — Paved development loops and production-grounded validation

4 documented initiatives · 2 with public metrics · 4 lifecycle scopes

Netflix emphasizes specialist paved paths. Its performance-agent workflow reads production profiles, checks out the exact commit running in production, matches against a markdown pattern catalog, runs unit tests and canary deployments, and sends a human-approved code review — with CPU opportunities of 0.5–4.6% found across seven services. Metaflow's spin lets a coding agent build and test ML workflow steps incrementally. The oci-agent case, whose estimate reached only 25% of a polished baseline, shows why specialist paved paths and actor-critic checking matter.

Airbnb — One landmark migration, newer platform signals

7 documented initiatives · 2 with public metrics · 7 lifecycle scopes

Airbnb's strongest primary case is the Enzyme-to-React-Testing-Library migration: work broken into per-file parallel steps with retries, richer context, and feedback loops migrated 75% of target files in four hours during one stage and completed nearly 3.5K files in six weeks — versus 1.5 engineer-years estimated manually. Newer AirDev isolated remote environments, AirChat agentic coding, async agent execution, and CI-for-agents details come from job listings and secondary reporting.

Anthropic — Clearest reusable harness design guidance

12 documented initiatives · 1 with public metrics · 9 lifecycle scopes

Anthropic publishes the most reusable harness playbook: tool/ACI design, context management, and long-running work split between an initializer agent and coding agents that leave durable artifacts between sessions. Sandboxing for Claude Code — OS-level filesystem and network isolation — reduced permission prompts by 84% while shrinking the attack surface, a rare public metric for governance rather than throughput.

A seven-layer harness reference architecture

Synthesizing the 51 initiatives, one architecture recurs. Each layer is grounded in documented practice.

1
Intake, spec and risk classification

Classify work before an agent touches it. Zalando's risk bot routes 33% of PRs to auto-approval; GitHub scopes its coding agent to low-to-medium complexity tasks in well-tested codebases.

2
Context plane

Repository instructions, retrieval, ownership and dependency data. Netflix's performance agent checks out the exact production commit and reads a markdown pattern catalog; Metaflow uses CLAUDE.md instructions; Airbnb's migration succeeded by adding richer context on retries.

3
Planner / orchestrator

Bounded tasks, retries, budgets, durable state. Airbnb decomposed 3.5K files into per-file parallel steps; Anthropic separates an initializer agent from coding agents with artifacts between sessions; the Manage–Execute–Audit loop keeps task state outside the growing context.

4
Tool / ACI layer

Clear semantics, minimal surface. Stripe exposes a curated subset of 400+ MCP tools per agent because over-tooling degrades performance; Anthropic's guidance centres on agent-computer interface design.

5
Isolated runtime

Ephemeral sandbox or worktree, scoped filesystem and network. Anthropic's OS-level isolation cut permission prompts 84%; GitHub's agent boots a VM per task with limited internet; Microsoft's THINKINGBOX isolates MCP tool sessions; Airbnb's AirDev provides remote isolated environments.

6
Verification plane

Tests, static analysis, telemetry, canaries, independent evaluators, human approval. PerfInsights' dual LLM validation cut false positives from 80%+ to the low teens; Netflix runs unit tests plus canary CPU comparison plus human review; LangChain's self-verification loop and METR's advisor-actor-rater scaffold show the actor-critic pattern.

7
Operations and governance

Traces, cost and latency, audit, secrets, policy, kill switches, continuous evaluation. Uber's uReview benchmark tracks precision, recall, cost and latency per review; Google's Agent Quality Flywheel uses OTel traces, AutoRaters and failure clustering; Zalando's proxy centralizes access and cost.

Others and reusable patterns

Block — CI/CD Goose Secondary

Running in CI/CD pipeline

Block — Goose Primary

migrate legacy codebases, refactor and simplify complex logic, generate unit tests, streamline dependency upgrades, and speed up triage workflows

Block — Goose Janitor Primary

Goose recipes

Block — Headless Goose Secondary

Running in an automated background environment when an issue or PR is created and tagged

DreamX Team, Alibaba Group — Manage-Execute-Audit (MEA) loop Research

A manager to maintain the task state and determine the next subtask, a fresh-context executor to perform it, and a read-only auditor to verify the resulting environment state before the next round.

LongHorizon-Harness improves Qwen 3.7-Plus from 51.8% to 80.7% on WeaveBench, from 69.7% to 77.2% on Terminal-Bench 2.1, and from 2.8% to 8.3% on OSWorld 2.0.
GitHub — Copilot coding agent Secondary

Boots a VM, clones the repo, pushes commits to a draft pull request, and tags developers for review. Limited internet access to trusted destinations, CI/CD workflows require human approval before running, and existing branch protections apply.

GitHub — GitHub Copilot agentic harness Secondary

orchestration layer, multi-step reasoning chains, external tools such as linters or test runners, context retrieval, and error recovery

GitHub and JetBrains — Codex integration in IDEs Primary

IDE workflows

Google — AI Operator Secondary

production incident response and autonomous mitigation

Google — Actus Secondary

physical execution control planes for safe, autonomous mitigation

Google — Agent Quality Flywheel Primary

GenAI evaluation service, User Simulator, AutoRaters, custom rubrics, Automatic Loss Analysis, OTel traces

0% to 96% of responses across all 15 cases
Google — Gemini CLI Secondary

Async generator loop, ModelRouterService with pluggable strategies, Scheduler state machine for tool execution, and hybrid loop detection via SHA-256 hashing and adaptive LLM self-checks.

Google — Investigation Dashboards (InvD) Secondary

generating dynamic, incident-specific single pane of glass interfaces to reduce manual data gathering during incidents

LangChain — Self-verification loops Secondary

PreCompletionChecklistMiddleware

66.5%
METR — advisor-actor-rater loop Research

1. Prompt the model to analyze the present situation and generate advice. 2. Have the model generate 6 possible actions it could take next (3 of which are generated with the advice in context, and 3 of which are not). 3. Tell the model to rate each of the 6 po

Performance comparable to what human baseliners can achieve in 2 hours
Microsoft — THINKINGBOX Research

isolated MCP-compatible tool sessions, complete execution traces, and outcome evaluation over terminal backend state.

Across proprietary and open-weight models, the strongest achieves 65.36% pass@1, but only 25.25% passˆ20.
Stripe — Minions Secondary

deterministic harness nodes for CI and linting, agentic nodes for implementation, a harness-enforced limit of two CI rounds before escalation to a human, and a curated subset of over 400 MCP tools exposed per agent

over a thousand pull requests per week

All 39 focal-company initiatives

CompanyInitiativeLifecycle scopePublic metricTierSource
ZalandoAutomated UI component library migrationMigration and Refactoringless than $40 for each code repository; accuracy of about 90%Primaryengineering.zalando.com
ZalandoChat UI and CLIcodingPrimaryengineering.zalando.com
ZalandoCLI toolmaintenance scripts for model access in the terminal and coding tasksPrimaryengineering.zalando.com
Zalandointelligent SRE assistant postmortem analysis pipelineOperations3x productivity boost in summariesPrimaryengineering.zalando.com
ZalandoLiteLLM API proxycoding2k MAU with just six small (2k CPU, 4 GB) podsPrimaryengineering.zalando.com
ZalandoLLM-as-a-judge search quality assurance frameworkEvaluating and assuring search quality at scale with multi-language support for pre-market launches and existing marketsPrimaryengineering.zalando.com
ZalandoMulti-stage LLM pipelineOperationsProcessed annual data analysis in under 24 hoursPrimaryengineering.zalando.com
ZalandoPython based migration toolMigrating partner facing applications from one of the UI component libraries to the other onePrimaryengineering.zalando.com
ZalandoRisk-based PR approval toolcode review33% of our PRs are low-risk and are auto-approved by the botPrimaryengineering.zalando.com
UberAutocovergenerating unit testsgenerating 5,000+ unit tests per monthSecondarynewsletter.pragmaticengineer.com
UberDragonCrawlTestingblocked ten high-priority bugs from impacting customers while saving thousands of developer hours and reducing test maintenance costsPrimaryuber.com
UberLLMCheckTesting and ValidationAchieved a 33.5% reduction in validated static analysis findings in four months (from 265 down to 176).Primaryuber.com
UberMinionships about 1,800 code changes a weekabout 1,800 code changes a week, and 95% of Uber's engineers use itSecondarysandeep.dev
UberPerfInsightsTesting and OptimizationReduced false positives from over 80% to the low teens.Primaryuber.com
UberSoftware Factory agentic offeringscode review, self-healing CI failures, completing E2E PRs with visual validation, triaging on-call alerts, debugging incweekly active users grew 7x, weekly agentic requests grew 9.4xSecondaryx.com
UberuReviewAI code review for all pull requestsHandles code review for all pull requestsSecondaryx.com
NetflixAI agent-assisted performance engineering workflowAutomated performance profiling, bottleneck identification, pattern/anti-pattern recognition, and code review generationCPU and latency savings observed in production; found across seven different services with savings between 0.5 to 4.6% of CPU cycles.Secondaryyoutu.be
NetflixMetaflow spinIncremental development, testing, and debugging of ML and AI workflow steps using AI coding agentsPrimarynetflixtechblog.com
Netflixoci-agentAnalysis and Evaluation25% of the baseline estimatePrimarynetflixtechblog.com
NetflixSpindevelopment and debuggingPrimarynetflixtechblog.com
Airbnb@generateMockGraphQL data mocking at scalePrimarymedium.com
AirbnbAirChatagentic coding and software development tasksSecondarymarkolukicic.substack.com
AirbnbAirDevproviding isolated remote development environments for concurrent agentic workSecondarymarkolukicic.substack.com
Airbnbasync agent execution platformbuild, validate, review, and ship high-quality softwareSecondarycareers.airbnb.com
AirbnbCI platform infrastructurecreate, build, test, review, and merge PRs for AI coding agentsSecondarycareers.airbnb.com
AirbnbLLM-driven code migrationTestingupdated nearly 3.5K React component test files from Enzyme to use React Testing Library (RTL) instead in just 6 weeksPrimarymedium.com
AirbnbLLM-driven migration pipelineupdating nearly 3.5K React component test files from Enzyme to use React Testing Library (RTL) insteadmigrated 75% of our target files in just four hoursPrimarymedium.com
AnthropicClaude Agent SDKcoding and environment setupPrimaryanthropic.com
AnthropicClaude CodecodingPrimaryclaude.com
AnthropicClaude Code auto modeAutomated coding tasks and tool execution while delegating permission approvals to model-based classifiersPrimaryanthropic.com
AnthropicClaude Code on the webcoding and execution verificationPrimaryanthropic.com
AnthropicClaude CoworkCollaborative developer workflow and productivityPrimaryanthropic.com
AnthropicCode ReviewCode review for pull requestsPrimaryanthropic.com
Anthropiccoding agentto resolve SWE-bench tasks, which involve edits to many files based on a task descriptionPrimaryanthropic.com
AnthropicCoding Agent for SWE-benchcoding and testingPrimaryanthropic.com
AnthropicComputer use reference implementationexecutionPrimaryanthropic.com
AnthropicFilesystem isolationcoding and execution verificationPrimaryanthropic.com
AnthropicNetwork isolationcoding and execution verificationPrimaryanthropic.com
AnthropicSandboxing for Claude Codecoding and execution verification84% reduction in permission promptsPrimaryanthropic.com

Risks and evidence quality

Public counts are non-comparable: companies disclose different slices, definitions and time windows. Missing metrics are not zeros, and initiative counts do not measure maturity — Anthropic documents 12 initiatives with one metric; Uber documents 7 with 7.
First-party engineering posts are prioritized (Primary); job descriptions, talks, newsletters and social posts are labeled Secondary signals; benchmark papers are Research. Uber's Software Factory growth figures and Airbnb's AirDev/AirChat details are Secondary.
Failure modes documented in the rows: self-assessment bias (LangChain), state drift over long horizons (Manage–Execute–Audit), one-off success masking unreliability (THINKINGBOX's 65% pass@1 versus 25% pass^20), and generalist agents underperforming paved paths (Netflix oci-agent).

A 90-day adoption roadmap

Days 1–30 · Foundations

  • Stand up a model proxy for access, cost and audit (Zalando pattern).
  • Write repository instruction files and a pattern catalog.
  • Classify task risk tiers; pick two bounded pilots — a mechanical migration and test generation.

Days 31–60 · Harness

  • Isolated runtime: ephemeral sandbox or worktree, scoped filesystem and network.
  • Minimal curated tool surface; deterministic steps stay deterministic.
  • Verification plane: tests plus an independent evaluator; cap retries (Stripe's two-CI-round limit) before human escalation.

Days 61–90 · Operate

  • Traces, cost and latency dashboards, kill switches, secrets policy.
  • Continuous evaluation on real historical tasks (uReview-style benchmark).
  • Expand from pilots to background agents on low-risk queues with human review gates.
KPIs to track from day one: task success rate · review acceptance rate · escaped defects · regression rate · cycle time · cost per accepted change · retries per task · permission prompts per session · human interventions per task · rollback rate

Method: 51 publicly documented agentic/harness engineering initiatives compiled to 2 September 2026 — 39 at Zalando, Uber, Netflix, Airbnb and Anthropic, 12 from other organizations — each with lifecycle scope, harness detail, public metric where disclosed, source tier and URL. Counts are documented public examples, not internal system totals; metrics use each company's own definitions and windows and are not cross-comparable. Some initiative descriptions were shortened for space; lifecycle categories in the matrix were assigned from each row's stated scope.

This report was generated automatically by Keenable SELECT at a user's request, from publicly available web sources linked herein. Keenable does not review, verify, or endorse its contents and makes no representation as to accuracy, completeness, or timeliness; AI-based extraction may contain errors. Nothing in this report is investment, legal, financial, or other professional advice. All trademarks and referenced content remain the property of their respective owners; no affiliation or endorsement is implied. To report an error, rights concern, or request removal: legal@keenable.ai.

Keenable SELECTAsk your own question
Made with Keenable SELECT