The Agentic Development Spiral

Forget the Agentic loop, think agentic spiral.

This article walks through an interesting pattern in Agentic AI development and explores its implications for the agentic enterprise and the people making it happen. It was inspired by a mathematician named Euler, who’s theory and equation shows how each turn around a circle actually describes a spiral, when plotted over time. This spiral casts a perpendicular shadow of Sine and Cosine waves on adjacent planes (see diagram) with the projected circle as the base.

This visually decomposes the complex unit circle into its real (cosine) and imaginary (sine) components across three dimensions to show how they oscillate over time. It is considered the most beautiful equation in mathematics, and calls to mind the adage of how it’s impossible the same river twice, (because the waters you crossed the first time has moved on downstream).

But what has this got to do with agentic AI development, or emergning patterns?

How does this apply to Agentic AI?

Euler’s formula visualisation provides a great metaphor for describing agentic development lifecycle and its evolution. As every AI Builder knows (or should know!), the Agentic AI development lifecycle typically spans the following stages or some version of it:

Plan / Define,
Design / Build,
Evaluate,
Deploy
Observe / Extend

You can imagine how an agent in development traverses each stage of the lifecycle.

Over time, as the agent gains more capabilities, it will encounter novel challenges over and above existing ones. These capability lifts and related challenges are analogous to the sine and cosine waves on the formula’s real and imaginary plane. This animation shows how the agentic loop or development lifecycle evolves over time, much like Euler’s theorem.

Key Observations :

The loop remains – each turn of the circle is a full cycle back to the start.
Both shadows never go away they are sine and cosine of the same angle around the agentic loop – ship the new capability and you’ll get the challenges it brings
The challenges lag by a quarter turn – if you think about it the challenges are introduced by the capabilities – e.g. think of voice creating challenges around accents
Cone shaped blast radius (I prefer to think of this as splash zone) – th increasing with turn – i.e. net capability compounds but the impact of challenges also expands.

Lets Take It Step by Step

Lets take a closer look at the capability opportunities and related challenges of each stage of the agentic development spiral

GENERATION 1 – ASSISTED

A model in the conversation. Humans still do the work; the model shortens it.

EMERGING CAPABILITY & OPPORTUNITY

Natural-language surface over systems people already use
Drafting, summarising, classifying at near-zero marginal cost
Fast to demo — days from idea to something convincing

NOVEL COMPLEXITY & CONSTRAINT

Output can’t be verified — no source, no notion of truth
Prompt fragility: behaviour shifts with wording and model version
Value stays trapped in individual habit; nothing accrues to the org

ARCHITECTURAL CONSIDERATIONS

Treat prompts as versioned, tested artefacts — not config strings pasted into a field. Decide where the line sits between the model’s output and the system of record before anyone ships.

GENERATION 2 · GROUNDED

Retrieval arrives. The model answers about your business, not the internet’s.

EMERGING CAPABILITY & OPPORTUNITY

Answers anchored in enterprise knowledge, with citations
Domain specificity without retraining anything
Content you already own becomes a live asset

NOVEL COMPLEXITY & CONSTRAINT

Retrieval quality becomes the product — the model is now the easy part
Stale, duplicated and contradictory sources surface as confident answers
Permissions leak through the index: the retriever inherits every ACL you forgot to model

ARCHITECTURAL CONSIDERATIONS

Enforce access control at retrieval time, per identity — never by filtering results afterwards. Content lifecycle, ownership and freshness stop being a knowledge-management problem and become an architectural one.

GENERATION 3 · ACTIONS

Tools and writes. The agent stops describing work and starts doing it.

EMERGING CAPABILITY & OPPORTUNITY

Reads and writes to systems of record — real transactions, real records
Task completion becomes measurable, so value becomes attributable
Work moves from the chat window into the workflow

NOVEL COMPLEXITY & CONSTRAINT

Side effects are irreversible; a retry is not free
Non-idempotent actions, partial failures, half-finished transactions
Authorisation is now per tool, per caller — and blast radius scales with the toolset
“What did it actually do?” becomes an audit question with a regulator attached

ARCHITECTURAL CONSIDERATIONS

Every tool needs its own authorisation, an idempotency key, and an audit record — no shared service account. Budget for the reversal path, not just the happy path: compensating actions, dry-run modes, and approval gates on anything you can’t undo.

GENERATION 4 · AUTONOMOUS

Planning, memory, self-correction. It runs while nobody is watching.

EMERGING CAPABILITY & OPPORTUNITY

Goal decomposition across many steps and many systems
Memory across sessions — context stops being re-explained every time
Recovers from its own errors; handles the long tail you never specified

NOVEL COMPLEXITY & CONSTRAINT

Non-determinism defeats conventional testing — the same input, two different runs
Cost and latency vary per execution; unit economics become a distribution, not a number
Memory drifts, carrying stale assumptions forward as fact
Loops can run away — spend, steps and time all unbounded by default

ARCHITECTURAL CONSIDERATIONS

Move from assertion-based tests to evaluation: score distributions, tolerances, regression sets, and a golden path you re-run on every model change. Memory needs a schema, a TTL and a correction mechanism. Hard ceilings on spend, step count and wall-clock are runtime requirements, not ops hygiene.

GENERATION 5 · ORCHESTRATED

Agents calling agents — across teams, domains, and vendors you don’t control.

EMERGING CAPABILITY & OPPORTUNITY

Specialisation and parallelism — narrow agents outperform one general one
Cross-domain workflows that no single team owns end to end
Composition: capability assembled, not rebuilt

NOVEL COMPLEXITY & CONSTRAINT

Coordination overhead can exceed the gain — the classic distributed-systems tax, now non-deterministic
Attribution across hops: which agent caused the outcome, and on whose authority?
Trust boundaries with agents you don’t own; every hop is potentially hostile input
Failures cascade and are hard to localise; behaviour emerges that nobody specified

ARCHITECTURAL CONSIDERATIONS

Observability must be causal and end-to-end — one trace per goal, not one log per agent. Agents need identity, explicit contracts, and delegated (not inherited) authority. Assume any inbound message may be adversarial, including from your own agents: validate at every boundary.

GENERATION 6 · GOVERNED ESTATE

Agents as an organisational asset class — discoverable, owned, governed, retired.

EMERGING CAPABILITY & OPPORTUNITY

Reusable, discoverable agents: leverage compounds instead of being rebuilt per team
Policy applied once, inherited everywhere
The estate becomes plannable — capacity, cost and risk modelled like any other platform

NOVEL COMPLEXITY & CONSTRAINT

Version and lifecycle sprawl: hundreds of near-duplicate agents, unclear owners
Accountability and regulatory evidence — who is answerable for a decision the estate made?
Model supply-chain risk: upstream changes ripple through everything at once
Role and workforce change becomes the real constraint, not the technology

WHAT THIS FORCES IN THE ARCHITECTURE

A registry is non-negotiable: ownership, version, dependencies and blast radius per agent. Policy as code in the runtime path, not a document. Design the off-ramp on day one — deprecation, migration and decommissioning are the part everyone skips and everyone pays for.