The End of Software Engineering and the Rise of Agentic Engineering

Agentic Engineering

Under stated assumptions there is a structure-preserving interpretation of the four pillars of agent externalization, memory, skills, protocols and harness, in the ArchAgents Architecture triple (G, Know, Phi); three production systems, ContextFS, AgentHero and CatDB, realise it on identifiable fragments, and the series states where it fails.

4 parts + synthesisContextFS · AgentHero · CatDB

The papers

Each part fixes one pillar against one production system at a pinned commit, and states which conditions hold, which hold on a restricted fragment, and which fail.

Coalgebraic Memory paper preview
Memory · ContextFS

Coalgebraic Memory

42 pages

Agent memory is often called coalgebraic state, and the description rarely goes past the letters. We fix a functor, a set of laws, and a way of...

Operadic Skill Composition paper preview
Skills · AgentHero

Operadic Skill Composition

38 pages

The claim that skills are operations of a coloured operad is testable only after its colours, operations, substitution, and semantics are specified....

Typed Protocol Wiring paper preview
Protocols · CatDB

Typed Protocol Wiring

35 pages

The categorical reading of agent architecture assigns the protocol layer to the syntactic wiring G of an architecture triple (G,Know,Phi). We make it...

Harness Architecture paper preview
Harness · AgentHero harness runtime

Harness Architecture

42 pages

An agent harness surrounds a language model with tools, memory, wiring and policy. Banu proposes that this layer is already described by the...

How the series composes

The four parts are read in order: each layer’s interpretation is stated using the definitions the previous layer exports, and each adds one property while carrying forward the failure modes already found.

How the series composesMemory composes into Skills, Skills composes into Protocols, and Protocols composes into the Harness, each layer building on the one before it.MemorySkillsProtocolsHarness
Memory

Lineage runs on a single time axis, so ancestry is fixed once recorded and later derivations cannot rewrite it.

Skills

Composition is substitution-closed on a restricted colour set, well founded whenever sibling output keys stay disjoint.

Protocols

Every port is typed, and every way of declining is a distinct, observable outcome rather than a silent failure.

Harness

Certificates carry a scope and a decidable replay check, so what a run claims can be checked against a receipt.

Evidence

Every claim is checked against one production system at a pinned commit, not against a reference implementation built to satisfy the theory.

ContextFS at commit a93035d
AgentHero at commit 1c3ad24
CatDB at commit 7cc9341a

Browse the code-evidence layer for the file and line references behind every checked claim.