AGENTIC SDLC
Engineering the Workflow Agentic development is not "ask AI to build the app". It is the deliberate design of context, constraints, permissions, feedback loops, checkpoints, independent verification, escalation rules and ownership across the whole lifecycle, without surrendering engineering judgement.
THE LIFECYCLE AS A CONNECTED SYSTEM · SELECT A STAGE
DISCOVERY REQUIREMENTS RESEARCH ARCHITECTURE PLANNING CODING TESTING REVIEW DEBUGGING SECURITY DOCUMENTATION DEPLOYMENT MONITORING MAINTENANCE RETROSPECTIVES
Discovery Framing the problem before any agent sees it. Agents help survey prior art and competitors; the decision of what is worth building stays human.
WHO HOLDS RESPONSIBILITY HERE
HUMAN
IMPLEMENTATION AGENT
REVIEW AGENT
RESEARCH AGENT
TESTING / VALIDATION
CI/CD & PRODUCTION
WHERE AGENTS FAIL The workflow exists because these failure modes are predictable. Each one is a terminated path in the lattice: observed, named, and designed around.
APIs, flags and behaviours that never existed, delivered with total certainty. ⟲ MITIGATION Verify against real sources: documentation, code and running systems, never against the agent's memory. Claims without evidence are leads, not facts. ⟲ Flip for mitigation Structures that satisfy the prompt but ignore the system around them. ⟲ MITIGATION Adversarial architecture review before any code: a second agent, or the human, attacks the proposal until it accounts for the whole system. ⟲ Flip for mitigation Improving the file at the expense of the codebase. ⟲ MITIGATION Explicit scope in every task, plus a reviewer instructed to read beyond the diff and check the consumers of what changed. ⟲ Flip for mitigation Long sessions silently forget earlier constraints. ⟲ MITIGATION Short, scoped sessions with deliberate handover documents. Compress context on purpose before the agent loses it by accident. ⟲ Flip for mitigation Rewriting working code because rewriting is easier than reading. ⟲ MITIGATION Write permissions on named files only. Everything else is read-only by default. Rewriting must be requested, never assumed. ⟲ Flip for mitigation "Tests pass" because the tests were weakened until they passed. ⟲ MITIGATION Independent validation the author agent cannot touch: separate test runs, protected test files, and a reviewer who diffs the tests too. ⟲ Flip for mitigation Ten tool calls where one grep would do: cost, noise and latency. ⟲ MITIGATION Observe traces regularly and prune the workflow. Give agents cheaper deterministic shortcuts and instructions on when to use them. ⟲ Flip for mitigation Changes that work locally and break a consumer nobody mentioned. ⟲ MITIGATION Repository onboarding docs and dependency maps as standing context, so every session knows what depends on what before it edits. ⟲ Flip for mitigation Secrets in logs, permissive defaults, injection-friendly string building. ⟲ MITIGATION Deterministic scanning plus human review on every security-relevant change. Agent self-assessment is never accepted as evidence. ⟲ Flip for mitigation The user is wrong and the agent politely proceeds anyway. ⟲ MITIGATION An explicit challenge step whose only job is disagreement: an agent prompted to find what is wrong, not to be helpful. ⟲ Flip for mitigation