Disciplinary claim · Historical and empirical evidence
The End of Software Engineering
Software engineering is being redefined as agentic engineering. The decisive change is a new unit of work: engineers specify objectives and...
A five-part research series
The End of Software Engineering and the Rise of Agentic Engineering
Software engineering is being redefined as agentic engineering. Direct authorship is giving way to specifications, contracts, orchestrated agent activity, and verification of machine-produced artifacts. Engineering judgment becomes the control layer.
The papers
The argument moves from the disciplinary claim to system anatomy, the working loop, multi-agent coordination, and the context and data substrate.
Primitives and composition · Four first-party systems
The Anatomy of Agentic Engineering
Software is increasingly produced by directing programs that write, run, and repair code, but the vocabulary for what is built has not settled. This...
Execution and verification · Two code bases and one hosted product
Plan, Act, Verify, Repeat
Agentic engineering runs on a loop: state an intended change, act on it, verify the result, and decide whether to stop, repeat, or escalate. We ask...
Multi-agent coordination · Elixir and Rust runtimes
Agent Teams
Practitioners who outgrow one agent face three questions: when several agents are the right answer, how to arrange them, and what the arrangement...
Context, memory, and data · Four first-party systems
Context, Memory, and Governed Data
A context window is a bounded token sequence whose accuracy falls with length and varies with position, and an agent's reliable task length is...
How the series composes
Each paper owns one layer of the argument. Later papers apply the earlier definitions instead of silently redefining them.
Claim
Initiator, action count, and verification placement distinguish agentic execution from assisted programming and fixed workflows.
Primitives
Nine operational primitives and five composition operators define the system being discussed.
Loop
Planning, action, and verification repeat behind an independent gate and an explicit iteration bound.
Teams
Multiple loops compose through contracts, ownership rules, serialized resources, and integration review.
Substrate
Context, persistent memory, provenance, and governed data bound what those agents can know and safely return.
Evidence
Published studies, standards, benchmark audits, and first-party system evidence support different kinds of claims. Vendor results are identified as vendor results; source inspection establishes that a mechanism exists, not how often it succeeds in use.
Randomized productivity studies across distinct task classes
Independent benchmark and answer-key audits
Four first-party code bases inspected at pinned commits
Inspect the manuscripts, source index, claim ledgers, and review record behind the published pages.