Read the online version
The web edition provides the complete content without a form.
- Clickable contents and chapter anchors.
- Factual revision dated 6 September 2026.
White paper no. 3 · Revised 6 September 2026
Enterprise agentic processes and tomorrow's interfaces
When AI executes a process instead of producing an isolated answer, the organisation must control decisions, authorisations, checks and resumptions through a shared interface.
- Distinguish a nondeterministic skill from a workflow whose orchestration is coded.
- Calculate reliability with conditional probabilities and read each benchmark within its own scope.
- Adapt idempotence, validation, evaluation and authorisation practices to business risk.
- Design a control room and select a usage tier without treating autonomy as an end in itself.
What you will find inside
- Skill and agentic workflow: two objects that everyday vocabulary blends together, their respective promises, and the operations where model judgement provides value that must be evaluated.
- Conditional reliability: the independence calculation, its bounds without a dependence assumption, the separation of repeated trials from workflow stages, and the distinct results from Toolathlon and AutomationBench.
- Four software engineering practices: declared idempotence, machine validation, evaluations built upfront, human decisions and audit trail written into code, adapted to context rather than treated as a universal architecture.
- The control room: its four properties (decision queue, register of authorisations, append-only log, explicit resumption path) and the maturity test that sums them up in one question.
- The three tiers of use: assisted, supervised, bounded autonomy, and the five cumulative conditions to meet before reaching the last one.
- The grid of eleven disciplines: each software engineering discipline used as a design analogy for the agentic process, with an example outside code.
- The first deliverable: a map produced without buying a tool, its indicative duration for a simple case and four selection criteria aligned with at least monthly use.
Key figures
0.5
the minimum trigger rate a skill must reach over three runs to be accepted by the official test procedure.
Source: Agent Skills open standard, consulted 30 August 2026
36 % / 86 %
the rounded independent-model results for twenty stages against three at 95% per stage; the bounds without assumptions are 0–95% and 85–95%.
Source: White paper illustration under an independence assumption
73.1 %
success across three repeated trials of the same Toolathlon task, against 52.4% under independence; these trials are not workflow stages.
Source: Anthropic, System Card Claude Opus 5, 24 July 2026
26.0 %
full completion by the best reported model on AutomationBench scenarios spanning 47 applications, against under 10% in the April study.
Source: AutomationBench (Zapier), July 2026 reading
0 out of 240
unsafe tool actions observed with the metadata policy barrier across 240 NetInjectBench cases, with measured utility of 99.17% to 100%.
Source: NetInjectBench, 11 July 2026
21 %
readiness on the business-process dimension in Deloitte's sample of 501 executives from organisations already running agentic pilots.
Source: Deloitte, 11 August 2026 (501 executives)
Access white paper no. 3
The full text and PDF provide two formats for the same revised edition.
- Immediate online reading organised by chapter and anchor.
- A PDF copy available directly or by email request.
Request the PDF by email
The offline copy can be requested in French or English through a pre-filled message.
- French or English can be specified in the message.
- Direct English download available below.
Or download the EN PDF directly (no email) →
GDPR compliant. One-click unsubscribe. No data shared with third parties.
You can also write directly to contact@paulantoinetual.fr.
Frequently asked questions
- What is the difference between a skill and an agentic workflow?
-
A skill entrusts reusable know-how to the model's judgement, while a workflow puts the sequence into versioned, testable code.
- The skill leaves its triggering and the exact form of successive outputs open.
- The workflow fixes loops, filters, conditions and retries without making its model calls deterministic.
- The Claude Code documentation therefore distinguishes the “instructions” reused by a skill from “the orchestration itself” reused by a workflow.
- Why does reliability fall when nondeterministic steps are added?
-
End-to-end reliability depends on conditional probabilities between stages, and multiplying rates is valid only under an explicit independence assumption.
- Under independence and at 95% per stage, twenty stages give 35.8% and three stages give 85.7%.
- Without a dependence assumption, the bounds are 0% to 95% for twenty events and 85% to 95% for three.
- Toolathlon measures three repeated trials of the same task; its 73.1% rate does not represent three workflow stages.
- A validation or retry can improve overall reliability even when it adds a stage to the system.
- Do I need a tool or a budget to start?
-
A simple, already documented process can begin with a map produced without buying a tool, with a duration that depends on its complexity and participants.
- List the stages, their inputs, decisions and effects.
- Separate verifiable rules, judgement over content and committing actions.
- Estimate the nondeterministic calls to test and the controls to put into code.
- Treat half a day as an order of magnitude for a simple case rather than a guaranteed duration.
- How does this differ from white papers no. 1 and no. 2?
-
The three white papers move from maturity to industrialisation and then to the design and control of agentic processes.
- Volume no. 1 answers “where do I stand?”.
- Volume no. 2 answers “how do I industrialise?”.
- Volume no. 3 covers controls, interfaces and the conditions for autonomy in a process executed by agents.