ADLCThe Agentic Development Lifecycle
FIRST EDITION
THE SUCCESSOR TO SDLC

The AgenticDevelopmentLifecycle

Delegated execution under human supervision

AB CD
PRAVEEN GOVINDARAJ
Nothing opened yet

28 chapters · 5 appendices · 21 corrections logged · every claim graded A–D

The spine. SDLC assumed the bottleneck was producing code. ADLC assumes the bottleneck is verifying and directing it.

The argument in one figure

One price fell, and six goods that were resting on it fell together. Everything else in the book follows from that.

SIX GOODS THAT ARRIVED FREE — NONE OF THEM DESIGNED, NONE OF THEM AUDITED Scrutiny effort evidenced care Intent an author you could ask Pacing feedback outran change Trust boundary reading was never dangerous Homeostasis someone cleaned up eventually Apprenticeship juniors learned by doing cheap work THE COST OF PRODUCING CODE the price fell — and nobody was billed for what it had been buying EACH MUST NOW BE BOUGHT DELIBERATELY — WHICH IS WHAT A LIFECYCLE IS FOR Sensors blocking, drilled Ch 7 · 8 Specifications held-out, before Ch 8 · 12 Ordered controls by cost and speed Ch 8 · 12 Capability scoping by effect, not name Ch 11 · 13 Drift detection standing mandate Ch 4 · 12 Deliberate practice attempt first Ch 19 · 20 ADLC is the discipline of explicitly buying what the cost of writing code used to provide for free. Nobody removed six controls. One thing was withdrawn, and six goods were resting on it.
Figure 1 — The spine. The cost of producing code silently purchased six controls; when it collapsed, all six failed at once.

Four rings of control

Each ring exists because the one inside it saturates, and each outer ring makes the inner ones cheaper. This is an argument, not a taxonomy.

STRUCTURE reduce what must be known · Ch 10 GRAPH make what remains findable · Ch 9 HARNESS steer, then check · Ch 7–8 CONTEXT design the payload · Ch 6 MODEL not yours · not stable changes without notice ✗ saturates: the whole file scored 12.7% vs 18.0% for a window ✗ bounded by Ashby — a finite regulator, unbounded generator ~ reports the structure it is given; makes a tangle navigable, not small ▓ the only ring injection cannot cross READ OUTWARD Each ring exists because the one inside it has a ceiling. READ INWARD Each outer ring makes the inner ones cheaper. CROSSING ALL FOUR Verification & security Injection defeats every control that acts on what an agent believes, and cannot defeat one that bounds what it can do. A poisoned context cannot grant credentials it lacks.
Figure 2 — Context saturates, so the harness exists; the harness is bounded by Ashby, so structure exists. Verification and security cross all four.

The evidence, plotted

Five studies, one gradient. The two most rigorous sit at opposite ends — rigour does not determine sign, task selection does.

+56% +40% +26% −19% 0 +55.8% Peng et al. HTTP server, recruited devs ~+40% Heilman et al. 16,223 engineers, heaviest-use weeks +26.08% Cui et al. 4,867 devs, 3 RCTs SE 10.3% +21% Google RCT 96 engineers, loses significance −19% METR RCT 16 devs, own repos, 5 years' familiarity greenfield toy task ◄──── TASK REALISM · REPOSITORY FAMILIARITY ────► expert-owned brownfield The two most rigorously controlled studies sit at opposite ends. Rigour does not determine sign — task selection does. So the question is never "is there an RCT?" but "whose codebase was it?"
Figure 3 — Measured effect declines monotonically as the measurement approaches real work. No single point represents the line.

What teams actually built

Böckeler's quadrant, populated with adoption data from 2,853 repositories. The field invested in the weakest cell at roughly 29 to 1.

GUIDES — feedforward act before generation · prevent SENSORS — feedback act after · detect and self-correct COMPUTATIONAL deterministic ms–seconds reliable INFERENTIAL probabilistic slower semantically rich language servers · CLIs · scripts codemods · scaffolds · templates types that make bad states unrepresentable MCP 2.3–11.6% linters · type checkers · tests static analysis · fitness functions highest authority · never ages HOOKS — sensors inside the agent loop 0.9 – 3.2% least adopted of all eight mechanisms principles · rules · how-to guides reference documentation lowest authority · ages fastest CONTEXT FILES — CLAUDE.md, AGENTS.md 90.6% present in 2,586 of 2,853 repositories review agents · LLM-as-judge semantic analysis advisory adds review load; only blocking removes it Subagents 6.6–7.4% roughly 29 : 1
Figure 4 — Guides assert; sensors observe. Adoption concentrates in the cell with the least authority and the fastest depreciation.

The double loop

The inner loop changes the software. The outer loop changes what governs its construction. Their only connection is artefacts.

OUTER LOOP — THE STEERING CYCLE subject: THE HARNESS · days → quarters 7 · OBSERVE drift · drills · recurrence 8 · STEER install · and RETIRE 9 · GOVERN autonomy · incidents 10 · EVOLVE templates · platform machine-paced ▲ · human-paced ▼ INNER LOOP — THE DELEGATION CYCLE subject: THE SOFTWARE · minutes → hours INTENT held-out set GROUND context · graph DECOMPOSE bound the task GENERATE stochastic VERIFY by exception INTEGRATE deterministic variety is REDUCED ← ← ← · INJECTED · → → → COLLAPSED — the randomness comes to rest at the merge gate surprise at the output is a diagnostic about phases 1–3, never about the model ARTEFACTS — the loops' only connection retrieval traces · execution traces · sensor reports · escalations · recurrence counts no artefacts → no outer loop. Not a weak one — none.
Figure 5 — SDLC's feedback loop returns to requirements. ADLC's returns to the harness.

The autonomy ladder

A ceiling set by harnessability × blast radius, implemented as capabilities the agent does not hold.

EACH LEVEL IS A SET OF REVOKED CAPABILITIES — NOT A SET OF INSTRUCTIONS A0 Assisted human authors A1 Supervised approval each step A2 Delegated where the work is A3 Bounded-auto gated by sensors A4 Continuous runs unprompted Closed for ordinary feature work The behaviour harness is unsolved, so supervision cannot fall. Reachable today only via near-zero blast radius or mechanical acceptance. CEILING = HARNESSABILITY × BLAST RADIUS — NEVER ENTHUSIASM blast 0 blast 1 blast 2 harness high A4 A3 A1/A0 harness low A2 A1 A0 Blast radius vetoes. No amount of harness quality makes a severe-blast-radius change eligible for autonomy. You reduce the blast radius instead — or accept the ceiling.
Figure 6 — Blast radius vetoes: no amount of harness quality makes a severe-blast-radius change eligible for autonomy.

Chapters

Five parts, 28 chapters. Each defends one claim and carries graded evidence.

PART II

The Four Pillars

The engineering core: nested rings of control.
CHAPTER 06

Context Engineering: the Budget, Not the Bucket

The marginal token can be negative, so every inclusion is a costed choice — which means larger windows increase rather than reduce the need for the discipline, and partial observability must be attacked by findability and structure, not volume.

5159 words
CHAPTER 07

Harness Engineering I: Guides and Sensors

Guides assert, sensors observe — and the field has bought assertions over observations at roughly 29:1, concentrating investment in the quadrant's weakest and fastest-depreciating cell.

4506 words
CHAPTER 08

Harness Engineering II: The Steering Loop

Ashby's constraint means an unbounded generator cannot be comprehensively harnessed, so variety reduction is forced — and the only way to tell a working control from a decorative one is to deliberately produce the failure it exists to prevent.

5883 words
CHAPTER 09

Graph Engineering

A graph supplies information no model can be trained into rather than compensating for a capability — so it is the only feedforward guidance that does not age, and the architectural prose filling context files belongs in it.

4963 words
CHAPTER 10

Structural Engineering: Parnas and the Harnessable Codebase

Information hiding is the condition under which local inference is sound — but the elegant version overreaches, since nothing bounds what must be known about a reused module, and the mechanism awaits an experiment that would take a week.

5143 words
CHAPTER 11

Verification & Security

Injection defeats every control that operates on what an agent believes and cannot defeat one that bounds what it can do — so capability must be scoped by effect, never by name.

5628 words

Appendices

Reference material — patterns, templates, instrumentation, and the annotated bibliography including what this book rejected.

Reading paths

Four ways in, depending on why you are here.

Shipping software on Monday

21 → 13 → 14 → 26 → 8

Start with the anti-patterns — you will recognise yourself. Then the ladder, the metrics, the plan.

Intending to cite this

1 → 3 → 7 → 27

The boundary that makes the literature searchable, the evidence, the census, the open problems.

Sceptical any of this is new

2 → 4 → 5

The controlled comparison, the debt ledger, and what actually broke. If these three do not persuade you, nothing later will.

Enthusiastic and want slowing down

4 → 18 → 19 → 21

Persistent complexity, the perception inversion, what offloading costs, and the failure modes.