An agent is just a loop. Everything that matters is what you build around it.
Petr Stanislav · CTO & Co-Founder, Elin.ai · Prague
02
The speaker
Petr Stanislav
CTO & Co-Founder · Elin.ai · Prague
Building agentic loops in production — and the harness around them. This talk is what survived contact with real systems.
03
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 — the harness name rotates: the loop is the primitive, the tool is interchangeable
04
Act 0 — Where we are
The horizon is doubling every 7 months
20209 seconds
→
20234 minutes
→
late 202440 minutes
→
2025–26hours ← you are here
~12 h
at 50% reliability — roughly a working day, at coin-flip odds
~70 min
at 80% reliability — about an hour, if you need four-in-five
METR TH1.1, verified 2026-08-28 (Opus 4.6: 718.8 min at 50%, wide CI). Horizon = human-time of tasks done at 50% success — NOT autonomous runtime (MIT Tech Review: the "most misunderstood graph in AI").
05
Act 0 — Where we are
Every definition got shorter. None dropped the word loop.
2024 · Anthropic
"Agents are typically just LLMs using tools based on environmental feedback in a loop."
2025 · Moran / Willison
"Agents are models using tools in a loop."
2025 · Solomon Hykes
"An AI agent is an LLM wrecking its environment in a loop."
06Act 1 of 3
The Loop
the agentic loopmaker / checkerinner · middle · outercompound engineering
07
Act 1 — The loop
All the leverage lives on the verify edge
gather context→take action→verify→repeat
↩ the loop feeds back — each cycle starts fresh
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% severe vendor-internal — best with zero shared context between coder and reviewer.
Anthropic, "Getting started with loops" (2026) — "Loops that write code need loops that check it." · Cognition, "Multi-Agents: What's Actually Working" (2026-04)
08
Act 1 — The loop
The loop is trivial. The harness is the engineering.
1.6%
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.
"A decent model with a great harness beats a great model with a bad harness."
Addy Osmani, "Agent Harness Engineering" (2026)
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)
09
Act 1 — The loop
Autonomy is earned by verification, not granted by the task name
The generation–verification loop
Generatesmall diff, tight scope
↓
Verifychecks run, evidence collected
↓
Loosen the leashtrust grows, autonomy expands
Karpathy, "Software Is Changing (Again)" (2025)
The ceiling is set by risk — three questions
How fast will we know we're wrong?
How cleanly can we undo?
What would prove we're right?
10
Act 1 — The loop
Same primitive, four triggers
Turn-based
You verify each turn before the next one starts.
stop: you verify
Goal-based
Runs until an evaluator confirms completion — or hits a turn cap.
stop: evaluator or cap
Time-based
A clock or schedule ends the run.
stop: timer
Proactive
Event-driven, no human live — guardrails and escalation rules hold the line.
stop: guardrails
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)
11
Act 1 — The loop
Loops nest
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)
12
Act 1 — The loop
The bottleneck didn't disappear. It migrated.
+98%
PRs merged
+91%
review time
+154%
PR size
"We got faster cars, but the roads got more congested."
Addy Osmani, "The 80% Problem in Agentic Coding" (2026)
Faros AI/DORA, via Osmani (2026). Jevons paradox: "Teams are working more, not less" — Linear, How teams build (2026). 78% of Claude Code sessions are multi-file; ~20 autonomous actions before a human steps in.
13
Act 1 — The loop
Agents run the inner loop. Engineers own the outer loop.
Verdictthe production decision — ship, block, redirect
Answerability"if someone asks, I can explain why"
The boundary is evidence. The verdict is yours.
"Agents run the inner loop. Engineers own the outer loop." — Osmani, "Own the Outer Loop" (2026)
14
Act 1 — The loop
Add a fourth beat: the loop learns
plan→work→review→compound
↩ each cycle feeds the next — same loop, better system
Rule 01
Extract, don't repeat
Turn every edge case and bug fix into persistent harness rules.
Rule 02
The middle runs without you
plan → work → review should be boring and autonomous.
Rule 03
Document the thinking
The reasoning compounds across tasks, not the raw code.
Rule 04
Always ≥50% extraction
Invest half your iteration effort directly back into the system.
Kieran Klaassen, Every — Compound Engineering (2026): "The first three steps produce a feature. The fourth step produces a system that builds features better each time."
15
Act 1 — The loop
The live debate of 2026
Every · Compound Engineering
"Implementation is mostly solved. Judgment isn't."
Judgment is extractable — every verdict you make can be pre-loaded into the system as a rule.
VS
Osmani · Practical Loop Engineering
"You delegate the task, not the judgment."
Taste doesn't compile — the judgment stays with you.
16Act 2 of 3
The Graph
workflows vs agentsworkflow patternssingle-writer physicsplans as DAGsMCP · A2A
17
Act 2 — The graph
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)
18
Act 2 — The graph
Four shapes, no frameworks required
Prompt chaining
Fixed sequential steps with programmatic validation gates.
Routing
Classify input, dispatch to a specialized model or tool.
Parallelization
Scatter independent tasks or vote across attempts; merge.
Orchestrator–workers
Lead agent decomposes work and delegates to workers.
The fifth pattern — evaluator-optimizer — you already met in Act 1: it's the maker/checker loop.
"Actions carry implicit decisions, and conflicting decisions carry bad results." Parallel subagents are fragile — actions collide.
+1 day
June 13, 2025 · Anthropic
Multi-agent research system
+90.2% over single Opus · ~15× chat tokens internal eval
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)
20
Act 2 — The graph
Multi-agent works when writes stay single-threaded
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, 2026
Corroboration: Claude Code Agent Teams ship file locking + shared task lists — single-writer physics, productized.
21
Act 2 — The graph
You are the GIL of your AI agents
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.
"You are the GIL of your AI agents." — Addy Osmani, "The Orchestration Tax" (2026). Anti-patterns: permission laundering · summary substitution · fleet cosplay.
22
Act 2 — The graph
Codex and Claude Code converge on the same five pieces
automationsheartbeat
worktreesisolation
skillsintent debt paid down
plugins / connectorsreach
subagentsmaker/checker
Learn those five and your skills transfer across tools.
Schedules each task as its dependencies resolve — 3.6× faster.
The plan is a dependency graph, executed as fast as its edges allow.
LangChain, Plan-and-Execute Agents (2024)
24
Act 2 — The graph
The graph needs wiring. Exactly two directions.
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.
25Act 3 of 3
The Fleet
six wavesfleets in the wildrelease governancethe adoption barbell
26
Act 3 — The fleet
The waves overlap. They don't replace.
traditionalpre-2023
→
completions2023
→
chat2024
→
coding agentsH1 2025
→
agent clustersH2 2025
→
agent fleets2026 ← here
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)
27
Act 3 — The fleet
Green doesn't mean green
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 months internal, 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.
28
Act 3 — The fleet
Governing a fleet rhymes with the SDLC you already have
"Software teams didn't invent code review, CI/CD, and canary deployments because they enjoyed process."
Sierra, "Release governance" (2026)
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
intent.md
→
spec.md
→
plan.md
→
diff + tests
→
PR
→
incident
↩
new intent.md
Sierra, "Release governance: guardrails for agents at scale" (2026); Anthropic, "The AI-Native SDLC Playbook" (2026)
29
Act 3 — The fleet
Adoption is binary. Skill is a spectrum.
5–10%power users
~20%shallow use
~70%never use it
Varick — client rollouts
10% of seats burn 90% of tokens
vendor anecdote, corroborated pattern
Ronacher — bimodal poll
44% vs 44%
hand-written code: <10% vs >90%
Science — peer-reviewed
seniors gain · juniors: flat
gains concentrate in senior devs
"Adoption is binary, skill is a spectrum." — Varick (2026) · Ronacher poll (5,000 devs), via Osmani (2026) · Science (2025) · 88% of orgs adopted AI vs 6% seeing >5% EBIT impact (McKinsey 2025) · 95% of pilots show no measurable P&L impact (MIT NANDA 2025). For the 70%: the loop goes to the work, not the worker.
30
Act 3 — The fleet
What comes next
spec-firstwrite the spec before the code
→
spec-anchoredthe spec stays the reference
→
spec-as-sourceAWS Kiro · 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.
"What if the models don't stop getting good?"
Geoffrey Huntley
Chimera — yad.codes, "Building Chimera" (2026); Huntley auto-heal (2026-01); Altman via TBPN (2026-02); Hassabis via Wired (2026)
31
Act 3 — The fleet
The engineer's job moved up one level per act
write codewhere we started
→
design the loopact I
→
design the graphact II
→
own the judgmentact 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)
32
Thank you
Questions, disagreements, and war stories welcome.