01

talk --start

Agentic Engineering
Loops, Graphs, and What Comes Next

An agent is just a loop. Everything that matters is what you build around it.

02

act 0 — where we are

The Whole Talk in One Line

ralph.sh
$ while :; do cat PROMPT.md | claude-code ; done

Geoffrey Huntley, "Ralph Wiggum as a software engineer", July 2025 — archival: quoted as published; the CLI binary has since changed, do not modernize

03

act 0 — where we are

How Far Agents Have Come

9 s
2020
4 min
2023
40 min
late 2024
hours
2025–26
~12 h at 50% reliability
roughly a working day — at coin-flip odds
~70 min at 80% reliability
about an hour — if you need 4-in-5

Horizon = human-time of tasks done at 50% success — NOT autonomous runtime (MIT Tech Review: the "most misunderstood graph in AI"). Doubling ~7 months; ~89 days for models ≥2024.

METR TH1.1, verified 2026-08-28 (Opus 4.6: 718.8 min at 50%, wide CI); MIT Technology Review 2026-02

04

act 0 — where we are

What Is an Agent

2024

"Agents are typically just LLMs using tools based on environmental feedback in a loop."

Anthropic, 2024

2025

"Agents are models using tools in a loop"

Hannah Moran / Simon Willison, 2025

2025, honest version

"An AI agent is an LLM wrecking its environment in a loop."

Solomon Hykes

Every definition got shorter. None dropped the word loop.

05

Act I — The Loop

06

act 1 — the loop

The Agentic Loop

gather context
take action
verify
repeat
"Loops that write code need loops that check it."

Anthropic, "Getting started with loops" (2026)

Maker ≠ checker

"The agent that wrote the code isn't the one grading it." — Osmani

Fresh eyes win

Devin Review: avg 2 bugs/PR, ~58% severevendor-internal — works best with zero shared context between coder and reviewer.

This maker/checker shape returns in Act 2 as the evaluator-optimizer workflow pattern.

Cognition, "Multi-Agents: What's Actually Working" (2026-04) — self-reported; Claude Code docs (loop anatomy)

07

act 1 — the loop

1.6% Model, 98.4% Harness

98.4% operational harness

1.6% AI decision logic — the rest of Claude Code is the systems around the loop. Osmani independently lands on 10% model / 90% harness (Osmani, "The New Software Lifecycle").

"A decent model with a great harness beats a great model with a bad harness."

Addy Osmani, "Agent Harness Engineering" (2026)

Ralph

zero harness — a while-loop and a prompt file. It works.

Claude Code

maximal harness — 98.4% of the code is scaffolding around the same loop. It works better.

Rules, skills, and hooks are harness components — the previous talk covered how to build them.

arXiv 2604.14228, "Dive into Claude Code" (2026); Osmani, "The New Software Lifecycle" (2026)

08

act 1 — the loop

Verification Earns Autonomy

Autonomy is earned by verification — not granted by the task name.

The generation–verification loop

Fast generation + tight verification = reliability. Small diffs; loosen the leash only as verified trust grows.

Karpathy, "Software Is Changing (Again)" (2025)

One slider isn't enough

orchestration ↑ agency → L0 assist L2 scoped delegation L5 managed by exception

Osmani, "Agentic Autonomy Levels" (2026)

How fast will we know we're wrong?

How cleanly can we undo?

What would prove we're right?

09

act 1 — the loop

Four Kinds of Loops

Turn-based

trigger: your message
stop: every turn — you verify each one

Goal-based

trigger: /goal + explicit completion criteria
stop: evaluator confirms, or turn cap

Time-based

trigger: the clock — /loop 5m, /schedule
stop: timer or schedule ends

Proactive

trigger: an event — no human live
stop: guardrails and escalation rules

Loop = heartbeat, goal = hands. What loops don't buy you: vague goals and taste.

Anthropic, "Getting started with loops" (2026); independently confirmed by Osmani, "Practical Loop Engineering" (2026)

10

act 1 — the loop

Inner, Middle, Outer

outer — market feedbackweeks
middle — engineer steeringhours
inner — the agentminutes

gather → act → verify → repeat

"A product moves at the pace of its slowest loop."

Andrew Ng, The Batch (2026)

11

act 1 — the loop

The Bottleneck Migrated

+98%
PRs merged

+91%
review time

+154%
PR size

Faros AI/DORA, via Osmani, "The 80% Problem in Agentic Coding" (2026)

"We got faster cars, but the roads got more congested."

Addy Osmani

Jevons paradox: "Teams are working more, not less — AI has a Jevons paradox quality beyond token consumption."

Linear, How teams build (2026)

12

act 1 — the loop

Own the Outer Loop

"Agents run the inner loop. Engineers own the outer loop."

Addy Osmani, "Own the Outer Loop" (2026)

INNER · agent
capability — checks run
— evidence crosses →
OUTER · engineer
agency — a human decides
Quality
checks produce evidence
Verdict
the production decision
Answerability
"if someone asks, I can explain why"
13

act 1 — the loop

The Loop That Learns

plan
work
review
compound
"The first three steps produce a feature. The fourth step produces a system that builds features better each time."

Kieran Klaassen, Every — Compound Engineering (2026)

Extract, don't repeat

The middle runs without you
plan→work→review should be boring

Document the thinking
the reasoning compounds, not the code

Always ≥50% extraction

14

act 1 — the loop

The Judgment Debate

"Implementation is mostly solved. Judgment isn't."

Every, Compound Engineering

vs

"You delegate the task, not the judgment."

Addy Osmani, Practical Loop Engineering

15

Act II — The Graph

16

act 2 — the graph

Workflows vs Agents

The canonical axis: who decides the next step — code or the model?

Workflows

"LLMs and tools orchestrated through predefined code paths."

code decides the next step

Agents

"LLMs dynamically direct their own processes and tool usage."

the model decides the next step

"The most successful implementations weren't using complex frameworks... simple, composable patterns."

Anthropic, "Building Effective Agents" (2024)

17

act 2 — the graph

Four Workflow Patterns

Prompt chaining

stepgatestep

Routing

router
ABC

Parallelization

split
§§§
merge / vote

Orchestrator-workers

w1orchestratorw2
w3

The fifth pattern — evaluator-optimizer — you already met in Act 1: it's the maker/checker loop.

Anthropic, "Building Effective Agents" (2024) — pattern catalog

18

act 2 — the graph

One Day Apart

June 12, 2025 · Cognition

"Don't Build Multi-Agents"

"Actions carry implicit decisions, and conflicting decisions carry bad results."

+1
day

June 13, 2025 · Anthropic

Multi-agent research system

+90.2% over single Opusinternal breadth-first research eval
~15× chat tokensinternal breadth-first research eval

Both were right: read-heavy research parallelizes; write-heavy coding doesn't.

Cognition, "Don't Build Multi-Agents" (2025-06-12); Anthropic, "How we built our multi-agent research system" (2025-06-13)

19

act 2 — the graph

Single-Writer Physics

Multi-agent works when writes stay single-threaded and extra agents contribute intelligence, not actions. (Cognition, 2026)

Read-only scouts

many agents explore; one agent writes

Generator–verifier

one drafts, a fresh-context agent grades

Map-reduce-and-manage

decompose → parallel reads → one managed merge

"Unstructured swarms are mostly a distraction. The practical shape is map-reduce-and-manage."

Cognition, "Multi-Agents: What's Actually Working" (2026)

Corroboration: Claude Code Agent Teams ship file locking + shared task lists — single-writer physics, productized.

20

act 2 — the graph

You Are the GIL

"You are the GIL of your AI agents."

Addy Osmani, "The Orchestration Tax" (2026)

1 agentagent work — parallelizesyour judgment — serial
8 agents÷8your judgment — same width

"Spawning 8 agents doesn't speed up your judgment time. It just makes the queue deeper."

Anti-patterns: permission laundering · summary substitution · fleet cosplay

21

act 2 — the graph

Converging Primitives

Codex and Claude Code converge on the same five pieces. (Osmani, "Loop Engineering", 2026)

automations
heartbeat
worktrees
isolation
skills
intent debt paid down
plugins / connectors
reach
subagents
maker/checker

Framework mental models — labels, not reading material:

LangGraph
state machine
OpenAI Agents SDK
handoffs-as-tools
Google ADK
workflow primitives
MS Agent Framework
graph dataflow
Deep Agents
harness
22

act 2 — the graph

Plans Become Graphs

ReAct — a chain

thinkactthinkact

one LLM call per step — myopic, strictly serial

LLMCompiler — a DAG

planner
tasktasktask
join

streams a DAG of tasks; schedules each on dependency readiness — 3.6× speedup

This is where "graphs" literally means graphs: the plan is a dependency graph, executed as fast as its edges allow.

LangChain, Plan-and-Execute Agents (2024)

23

act 2 — the graph

MCP and A2A

agent↔  A2A — horizontal  ↔agent
↕ MCP — vertical ↕
tools · context

MCP

vertical: agent ↔ tool/context. Launched 2024; Linux Foundation 2025.

A2A

horizontal: agent ↔ agent — Agent Cards, tasks, artifacts. Announced 2025; Linux Foundation 2025.

MCP in practice — covered in the previous talk.

24

Act III — The Fleet

25

act 3 — the fleet

Six Waves, Three Tiers

traditional
pre-2023
completions
2023
chat
2024
coding agents
H1 2025
agent clusters
H2 2025
agent fleets
2026

Where fleets run today — three orchestration tiers:

In-process subagents

spawned inside one session — maker/checker, scouts

Local worktree orchestrators

parallel checkouts on your machine, one writer each

Cloud async

Copilot coding agent · Jules · Codex Web · Claude Code Web

You supervise agents like junior devs — scoped tasks, checked work, earned trust.

Steve Yegge, "Revenge of the Junior Developer" (2025); tiers: Addy Osmani, "The Code Agent Orchestra" (2026)

26

act 3 — the fleet

Fleets in the Wild

Cursor — Wilson Lin

Hundreds of agents on one project for a week: planners, sub-planners, workers — and a judge agent deciding what counts as done.

Wilson Lin / Cursor experiment, via Simon Willison, 2026

OpenAI — Harness team

~1M lines · ~1,500 PRs · zero hand-written lines in 5 monthsinternal, unaudited

Humans design the environment: design maps, quality grades, golden rules.

engineering.fyi, "Harness Engineering" (2026)

"Green doesn't mean green."

Addy Osmani, "Human judgment doesn't leave the software factory" (2026)

Agents game tests and silently drop features — demand evidence, not dashboards.

27

act 3 — the fleet

Governance Rhymes

Agent checks

"a linter for your agent"

Simulations

conversation-level gates before release

Merge approval

a Reviewer role for agent changes

Split-traffic canary

immutable snapshots → instant rollback

"Software teams didn't invent code review, CI/CD, and canary deployments because they enjoyed process."

Sierra, "Release governance" (2026)

intent.mdspec.mdplan.mddiff + testsPRincidentnew intent.md

The SDLC becomes a loop of committed artifacts — an incident re-enters as a new intent.md, with no human in the invocation path.

Sierra, "Release governance: guardrails for agents at scale" (2026); Anthropic, "The AI-Native SDLC Playbook" (2026)

28

act 3 — the fleet

The Adoption Barbell

"Adoption is binary, skill is a spectrum." (Varick, 2026)

5–10%power users
~20% shallow use
~70%never use it

Varick — client rollouts

5–10% power / 20% shallow / 70% never; 10% of seats burn 90% of the tokensvendor anecdote, corroborated pattern

Varick Agents, "AI Adoption is a Myth" (2026)

Ronacher — bimodal poll

44% write <10% of code manually vs 44% still >90% by hand

Armin Ronacher poll (5,000 devs), via Osmani, "The 80% Problem" (2026)

Science — peer-reviewed

Gains concentrate in senior devs; juniors show no measurable benefit

Science (2025), AI-assisted code study

88% of orgs adopted AI vs 6% seeing >5% EBIT impact (McKinsey 2025) · 95% of pilots: no measurable P&L impact (MIT NANDA)

McKinsey State of AI (2025); MIT NANDA, The GenAI Divide (2025)

For the 70%: the loop goes to the work, not the worker.

29

act 3 — the fleet

What Comes Next

spec-first
write the spec before the code
spec-anchored
the spec stays the reference
spec-as-source
AWS Kiro · GitHub Spec Kit · Tessl

Self-improving harnesses

Chimera: an agent framework built by agents (~144k lines). Huntley's software factory: found a bug, fixed it, deployed, verified — autonomous auto-heal (Jan 2026).

The economics tension

Altman: 2–3-person billion-dollar companies. Hassabis: 3–4× productivity should mean building more, not cutting people.

Chimera — yad.codes, "Building Chimera" (2026); Huntley auto-heal (2026-01); Altman via TBPN (2026-02); Hassabis via Wired (2026)

"What if the models don't stop getting good?"

Geoffrey Huntley

30

act 3 — the fleet

Judgment Stays Above

write code
where we started
design the loop
act I
design the graph
act II
own the judgment
act III

"Human judgment doesn't leave the software factory. It relocates."

Addy Osmani (2026)

"The loop keeps running. Human judgement stays above it."

Anthropic, "The AI-Native SDLC Playbook" (2026)