Autonomous ADLC
for FDEs

A development lifecycle in which agents help turn agreed requirements into working, verified and operated software, with explicit limits on execution and release authority. The main focus is agents developing software from specifications. The software can be an API, application, pipeline or agent. The guide also explains the additional evaluations needed when the product itself is an agent.

Yes. A worked recovery change connects customer intent, a PRD, functional and non-functional requirements, BDD examples, an agent task plan and acceptance evidence. The companion maps requirements to executable Python tests and keeps production integration, performance and operating checks visibly pending. The Gherkin file is a reading example; the bundled runner uses unittest.

Yes. Dedicated chapters cover skill packaging, package digests, metadata validation, deterministic helper tests, trigger and negative-trigger trials, adversarial inputs and promotion. The companion includes a latency-review skill and tested numerical helper.

Yes. The guide distinguishes advertised capabilities from authorization, explains the July 2026 MCP revision and A2A context, and designs RPC contracts around stable operations. The local companion contains a deliberately limited JSON-RPC dispatcher; it is not an MCP server or a network authentication implementation. The linked Agentic Protocols guide provides a separate local MCP track.

A code property graph combines program representations so an engineer can follow syntax, control flow and data flow. The guide explains a refund example, how to build a useful evidence packet and where static analysis has limitations.

A resettable environment supplies controlled tasks, an independent evaluator checks actual effects, and a bounded optimizer proposes candidates. The guide distinguishes the AgentGym research project from its own local teaching gym, then explains hill climbing, held-out acceptance and protected controls for recursive improvement.

The Python/SQLite companion runs locally without external packages, model keys or cloud credentials. Its exact extracted archive passed 54 tests and detected 10 deliberately broken controls. It demonstrates release approval, recovery after a lost response, concurrent budget reservations, skill validation, an evaluation gym and a bounded search. It includes 12 editable worksheets.

The companion uses local identities, editable SQLite stores and synthetic fixtures. It does not authenticate production users, execute untrusted code in a sandbox, deploy to a cloud or benchmark a live model. The guide identifies the separate adapters, protected services and acceptance evidence needed for those uses.

The edition was checked on 24 October 2026 against official company, standards and open-source project documentation. It distinguishes framework, harness and runtime responsibilities, product-specific availability, model migrations and the exceptions in Google's Gemini CLI to Antigravity CLI transition. Recheck the linked owner documentation before adoption.

Forward deployed engineers and application engineers building customer agents or coding-agent workflows. Basic Python, Git and HTTP familiarity helps. The early chapters establish the vocabulary before implementation, operations and improvement.

The PDF and companion are included with active paid Premium or approved complimentary guide access. Three complete sample pages are public. Referral-only online Premium does not include guide downloads. A lawfully downloaded copy may be retained for study under the publication notice; premium files may not be redistributed.

The learning paths and PDF link to Harness Engineering, Production AI Agents, Hard Deployments, concepts, questions, labs and work samples. Use a chapter for the lifecycle decision, then follow its linked lesson or exercise for implementation practice. Each resource keeps its displayed access level.

AI tools assisted research, writing, code and the original Gemini cover illustration. Technical figures are authored vector diagrams. Scenarios and numerical fixtures are labeled as synthetic. Automated checks and assistant review do not imply independent human editorial approval; the edition is prepared for publisher review.