In partnership with
8 bits for a Byte: Your organization is inside that second arrival right now. Whether you've named it or not.
This issue is about the teams that named it early – and exactly what they built while everyone else was still optimizing prompts. You'll get:
The organizational signal hiding in plain sight. LinkedIn sunset its APM program. Uber's AI costs rose 6x and their CFO is now in the room. These aren't tech stories. They're the leading indicators of career architecture in motion – and they tell you exactly where the wave is going before the benchmarks catch up.
The platform layer Uber built before broad rollout — MCP gateway, agent registry, CLI, telemetry. The infrastructure that separates scattered experiments from a governed agentic platform. If you don't have this yet, you don't have an Agentic Development Life Cycle (ADLC). You have technical debt forming in a new medium.
Why 70% of your engineers' daily workload can be delegated right now — and the one behavioral shift that sent developer NPS to an all-time high at Uber.
The spec isn't a planning document anymore. It's the machine-operable contract your agents execute against. Teams that haven't moved their PRDs into version control are not managing their agents. Their agents are managing their intentions.
The second arrival doesn't wait for consensus. Waves don't pause for late adopters.
Read this one front to back.
Book a Coaching 1:1 Call With Me
Walk into your next leadership meeting with a plan—not a pitch for more time. In one 30-minute session, I'll help you build a 10-page Strategic Implementation Framework tailored to your company'...
Let’s Get To It!
Welcome To AI Quick Bytes!
Bit 1: EVERY CIVILIZATION-ALTERING TECHNOLOGY ARRIVES TWICE — AND THE SECOND ARRIVAL IS HERE NOW
Every great technology arrives twice. The first time, it appears as a tool. The second time, it reorganizes everything the tool touches — the economy, the institution, the human role inside both.
The printing press was a tool. Then it restructured the Church, the university, and the nation-state. The factory was a tool. Then it restructured the family, the city, and the clock. The internet was a tool. Then it restructured commerce, attention, and political legitimacy.
The first arrival of AI gave us autocomplete, recommendation engines, and image filters. That was the tool phase. The second arrival is here now — and it is reorganizing the fundamental unit of how knowledge work gets done. Not the task. The lifecycle.
Uber's leadership named it directly: They declared AI one of six company-wide strategic shifts — moving from a "human-plus-early-AI" company to a generative AI-powered company. That framing matters. It isn't a tooling upgrade. It is a culture repositioning in a single earnings cycle.
Bit: Every technology's second arrival reorganizes everything its first arrival only touched.
Three Key Takeaways
The wave you're in determines whether AI is a feature or a foundation. The first arrival of AI gave your org tools. The second arrival is reorganizing how the lifecycle itself runs. If your AI strategy is still a list of tools, you're in the First Arrival mindset.
Strategic declarations are organizational signals. When Uber's CEO names AI as one of six company-wide strategic shifts — not a tech initiative, a strategic shift — that changes hiring, funding, and architecture decisions downstream.
The second arrival doesn't wait for consensus. Waves don't pause for late adopters. The organizations that named the Second Arrival early are already 18 months into infrastructure decisions that compound.
Action Summary: Name which arrival your organization is in.
Ask your leadership team this week: Is your AI strategy a tool list or a lifecycle transformation? The answer tells you which arrival you're responding to.
Audit your last three AI investments. Were they augmenting existing phases — or rebuilding how phases are executed? Augmentation is First Arrival thinking. Rebuilding is Second Arrival positioning.
Make a strategic declaration, not a project announcement. Uber didn't launch a "developer productivity tool." They declared a strategic shift. The declaration changes what's fundable, what's hirable, what's architecturally permissible.
A comprehensive guide for addressing the tax talent crisis
A labor shortage in tax is driving the need for a new skill set: one that blends technical tax knowledge with digital fluency.
Automation, AI and data-driven insights now define the role of tax professionals.
This new era of tax is not simply about adopting new tools, it’s about reshaping the skill set and mindset required to thrive in this field. Check out this guide for actionable insights into how to cultivate these skills with your team. See how advanced technologies can help bridge the tax tech gap to increase efficiency, ensure compliance, and drive better decision-making.
Bit 2:
Quote of the Week:
Your spec is your first line of defense against vibe coding — and your last line of defense against agentic drift.
Bit 3: THE SDLC WAS A MAGNIFICENT MACHINE — FOR THE WAVES THAT CAME BEFORE THIS ONE
The Software Development Lifecycle organized human beings into phases for fifty years. Plan, require, design, implement, test, deploy, maintain. Each phase created a handoff. Each handoff created a role. Each role created a specialization. Each specialization created a silo.
The SDLC was the right machine for the First and Second Waves — batch processing and client-server. Agile didn't eliminate it; Agile compressed its cycles. DevOps didn't replace it; DevOps automated it. Every previous wave adapted the SDLC. The Fourth Wave fills it.
An LLM-based agent can now plan, call tools, observe results, and iterate — autonomously executing what humans once executed sequentially. Uber built the platform infrastructure that makes this real: Michelangelo, their ML platform, has been extended into a full agentic layer — a central MCP gateway proxying internal and external tools with enforced authorization and telemetry, an agent registry with no-code and SDK builder tooling, and an AIFX CLI that provisions agent clients, installs MCPs, and connects engineers to production infrastructure. They didn't bolt agents onto an SDLC. They built the platform layer agents operate inside.
Bit: The First Wave gave us phases. The Fourth Wave gives us agents who execute them — and platforms that govern how.
Three Key Takeaways
Adapting the SDLC is not the same as replacing it. Agile and DevOps were adaptations. ADLC is a reorganization. Your teams will feel the difference: adaptation preserves roles; reorganization migrates them.
Platform architecture precedes agent deployment. Uber built MCP gateway, registry, and CLI before broad rollout — not after. Authorization, telemetry, and sandboxing are prerequisites. Your platform layer is the governance layer.
The silo is a symptom of the handoff — and the handoff is a symptom of human sequencing. When agents fill the phases, the coordination overhead that created the silo disappears with the handoff that required it.
Action Summary: Audit your platform layer before your next agent deployment.
Map your agent infrastructure this sprint. Do you have a central tool gateway with enforced authorization? An agent registry your teams can discover and reuse? If not, you don't have an agentic platform — you have scattered experiments.
Identify which SDLC handoffs in your org still require a human relay. Each one is a phase your agents aren't filling yet. Prioritize the highest-toil handoffs first — those are the ones agents can absorb with bounded risk.
Build telemetry before you scale. Uber's MCP gateway captures tool call telemetry from day one. You can't govern what you can't see. Instrument before you deploy at volume, not after you have an incident.
Learn how to code faster with AI in 5 mins a day
You're spending 40 hours a week writing code that AI could do in 10.
While you're grinding through pull requests, 200k+ engineers at OpenAI, Google & Meta are using AI to ship faster.
How?
The Code newsletter teaches them exactly which AI tools to use and how to use them.
Here's what you get:
AI coding techniques used by top engineers at top companies in just 5 mins a day
Tools and workflows that cut your coding time in half
Tech insights that keep you 6 months ahead
Sign up and get access to the Ultimate Claude code guide to ship 5X faster.
Bit 4: THE SHIFT FROM PAIR PROGRAMMING TO PEER PROGRAMMING — AND WHY IT CHANGES EVERYTHING
GitHub Copilot augmented development. Synchronous tab completion, an IDE chat window, faster authorship. Uber measured it: roughly a 10–15% bump in diff velocity. Meaningful — but still the First Arrival. Still a tool.
Peer programming replaced it. Background agents running fully autonomously, asynchronously, on multiple workloads at once. A developer prompts Uber's Minion platform, moves to the next problem, and receives a Slack notification with a PR link when the task completes. The developer reviews, approves, merges. The workflow isn't faster authorship — it's delegated execution.
The behavioral shift is profound. When Uber made agentic workflows available, 70% of workloads pushed into the system were toil tasks — library upgrades, dead code cleanup, bug fixes, migrations. High accuracy on bounded toil created a virtuous cycle: more success led to more delegation, which freed cognitive capacity for the work that grows the business. Developer NPS at Uber — the self-reported satisfaction score for engineering productivity — has never been higher. The inflection point correlates precisely with Minion's launch alongside Claude Sonnet and Opus. The gap between power users and casual users didn't close. It exploded.
Bit: The tool augments. The agent delegates. That is not a productivity improvement — it is a workflow reconstruction.
Three Key Takeaways
70% toil is not a surprising number — it's a diagnostic. When Uber's engineers voted with their prompts, they voted to offload the boring work. Your teams will do the same. The question is whether your infrastructure is ready to absorb it safely.
Peer programming multiplies through parallelism. A developer running three Minion tasks concurrently isn't 3x faster at coding — they're operating at a different level of abstraction. The ceiling on a single engineer's throughput just moved.
NPS is a leading indicator for adoption velocity. Uber's developer satisfaction score is the organizational signal that the workflow reconstruction is working. If your engineers don't feel meaningfully faster, your ADLC investment isn't compounding yet.
Action Summary: Build the infrastructure for delegated execution, not just assisted authorship.
Identify your highest-frequency toil tasks this week. Library upgrades, dead code cleanup, trivial bug fixes, migration scripts. These are the workloads with the highest agent accuracy and lowest blast radius. Start there.
Set up a background agent workflow for one bounded toil category before next sprint. Not a pilot. A production workflow with a real PR output, CI gates, and required review. Constrain scope; prove the loop; expand.
Measure developer NPS before and 60 days after deployment. Uber uses self-reported satisfaction as a leading signal. If your engineers don't report feeling meaningfully more capable, something in the workflow — prompt quality, tool access, review overhead — needs adjustment.
Bit 5: SPEC-DRIVEN DEVELOPMENT IS THE GOVERNANCE LAYER HIDING IN PLAIN SIGHT
Vague prompts produce vague agents. That sentence is the whole argument for spec-driven development — but most teams are still treating specs as planning artifacts rather than governance documents.
Spec-Driven Development treats the PRD as version-controlled code: reviewed, traceable, linked to executable acceptance tests that run in CI. GitHub's Spec Kit operationalizes this directly — requirements, motivations, and technical constraints defined before agents build, so the agent builds what was intended rather than what it inferred. Microsoft's AI-led SDLC framework names spec-driven development as the central mechanism for moving from a textual idea to requirements and scaffold, treating specs as first-class artifacts that evolve with the product.
Uber's AutoMigrate program is enterprise SDD in practice. Migration scope, transformation logic, and validation criteria live in YAML files in version control before a single PR is generated. Shephard — their campaign management platform — reads those specs to generate PRs, notify owners, refresh diffs, and track closure across hundreds of changes in a single migration. The spec isn't a planning document. It is the machine-operable contract the system executes against.
Bit: Your spec is your first line of defense against vibe coding — and your last line of defense against agent drift.
Three Key Takeaways
The spec is the governance document. If your PRDs aren't versioned, reviewed, and linked to executable tests in CI, you have ungoverned artifacts driving production behavior. You are not managing your agents — your agents are managing your intentions.
SDD unlocks multi-variant exploration without rewriting code. GitHub's Spec Kit makes this explicit: want to compare a Rust implementation to a Go implementation? Ask the agent to produce both against the same spec. The spec is the shared context that makes exploration cheap and convergence reliable.
Constitution.md is the org-level constraint you're probably missing. GitHub Spec Kit introduces a constitution document — non-negotiable principles that apply to every project before SDD iteration begins. Testing conventions, CLI-first requirements, security posture. Your organization has equivalent constraints. They should live in a file, not in someone's head.
Action Summary: Move your specs into version control this sprint.
Put your PRDs in-repo this sprint. Specs and prompts that aren't versioned, reviewed, and linked to executable tests are ungoverned artifacts driving production behavior. Version control is the minimum governance bar.
Convert acceptance criteria into CI-executable tests. Each PRD acceptance criterion should map to a test that runs automatically. This is the infrastructure that separates teams scaling ADLC from teams accumulating technical debt in a new medium.
Write a constitution.md for your most critical agent system. Define the non-negotiable constraints — security posture, testing requirements, least-privilege tool access — before any SDD iteration begins. The constitution is the governance document your agents can't violate.
Bit 6: Sunday Funnies
Bit 7: THE PROOF OF REORGANIZATION IS ALWAYS VISIBLE IN CAREER ARCHITECTURE FIRST
The Fourth Wave's most legible signal is never the technology announcement. It is what an organization does to its career architecture when it believes the wave is real.
LinkedIn is sunsetting its APM program — the training pipeline for product managers — and replacing it with an Associate Product Builder track that teaches coding, design, and PM as one unified discipline. They didn't restructure a job title. They restructured the pathway, the training, the performance criteria, and the hiring expectations simultaneously. Career architecture compounds forever. Tools change monthly.
Uber's signal is cultural and operational. Developer NPS has never been higher. AI costs have risen 6x since 2024 — a number their CFO now asks about directly, because the impact is no longer a rounding error. They tried top-down mandates to drive adoption. Those had some effect. What caused adoption to erupt was engineers sharing wins with engineers — peer promoters demonstrating what a Minion-generated PR looks like in practice.
This is how waves reorganize institutions. Not through mandates. Through demonstration compounding across the org until the new mode is obvious and the old mode is what people apologize for still doing.
Bit: Adaptation is not a competitive advantage. In a wave economy, it is the only form of survival.
Three Key Takeaways
Career architecture is the leading indicator — not the lagging one. LinkedIn restructured pathway, training, and performance criteria. Uber's adoption erupted through peer demonstration, not policy. Watch what your organization does to how it hires, calibrates, and recognizes engineers. That's where the wave becomes visible before the benchmarks catch up.
Peer demonstration outperforms top-down mandates for adoption. Uber tried directives. They helped — but what caused adoption to erupt was engineers showing engineers. Your most credible change agents are your engineers, not your directors.
Cost is the organizational forcing function. Uber's AI spend rose 6x in two years. That number now lives in the CFO conversation. If your AI investment isn't generating enough visible impact to justify that scrutiny, your program isn't compounding — it's accumulating cost with unclear returns.
Action Summary: Restructure your career architecture before your tools force you to.
Audit your performance review criteria this quarter. If AI fluency and agency aren't explicitly measured, you're calibrating on the wrong skills. Update one review rubric to include demonstrated agent orchestration and behavioral governance as performance signals.
Identify your three highest-leverage peer promoters. Find the engineers already using agents effectively and give them a formal platform to share wins. A 24-minute demo is more powerful than a six-slide strategy deck.
Get your CFO into the same conversation as your CTO. Uber's 6x AI cost increase created CFO scrutiny. That scrutiny is a governance forcing function. If your finance leadership isn't asking what AI is returning, your program isn't visible enough to defend — or to fund at the next level.
Bit 8: THE FLYWHEEL IS THE MEASUREMENT — AND YOUR COMPETITORS ARE ALREADY RUNNING IT
LinkedIn's measurement formula captures the ADLC flywheel precisely: (Experimentation volume × Experiment quality) ÷ Time from idea to launch. ADLC moves all three levers favorably. If your instrumentation can't show that, you can't manage it.
Every.to's compound engineering model proves the flywheel at its purest: one engineer maintains a 143,000-line codebase used 30,000 times daily. Bug fixes eliminate categories. Patterns become tools. The codebase compounds capability rather than compounds debt. Uber's flywheel runs on the same logic — AutoMigrate handles Java upgrades at scale, AutoCover generates 5,000 tests per month, and the maintenance agent auto-recovers 50% of failed builds. Each system runs the loop: build, evaluate, tune, redeploy — continuously. The teams that instrument this well compound. The teams that treat ADLC as "AI-assisted coding" will hit a ceiling they don't see coming.
Uber's CFO is now asking what AI is returning. Their AI costs rose 6x since 2024. That CFO scrutiny is the organizational signal that the flywheel is visible — and that the instrumentation needs to connect activity metrics to business outcomes. That is the unsolved problem for most teams: not whether the flywheel spins, but whether you can prove it to the people who fund it.
Bit: The flywheel runs on observability. Build the instruments before you need them.
Three Key Takeaways
Activity metrics are not business outcomes. Uber's leadership is explicit: diffs generated, PRs merged, and tests created are leading signals — not the number their CFO needs. Connect your ADLC instrumentation to feature velocity, time-to-experiment, and revenue impact before you're asked to.
The flywheel proof is in the compounding rate, not the starting number. Every.to's 143,000-line solo-maintained codebase didn't start there. It compounds to there. Your ADLC investment compounds the same way — but only if the evaluation, tuning, and redeployment loops are running continuously.
Cost scrutiny is a governance forcing function. When the CFO asks what AI is returning, that conversation either accelerates your program or audits it. Instrument business outcomes before you're in that room — not during it.
Action Summary: Connect your ADLC flywheel to the metrics your CFO and board understand.
Define one business outcome metric your ADLC investment should move. Feature velocity, time from design to experiment, revenue per engineer. Pick one. Instrument it. Present it alongside activity metrics in your next QBR.
Run the compound engineering audit on your codebase this quarter. Ask your engineering lead: does a bug fix eliminate a category of future bugs, or just the current one? The answer tells you whether your ADLC investment is building a flywheel or accumulating technical debt in a new medium.
Require behavioral metrics alongside operational metrics in every sprint review. Task success rates, escalation frequency, policy violations. Agent systems degrade as context changes and model behavior drifts. The teams that catch degradation early compound. The teams that catch it in a postmortem pay twice.
Until next time, take it one bit at a time!
Rob
P.S. Thanks for making it to the end—because this is where the future reveals itself.
