CtrlK
BlogDocsLog inGet started
Tessl Logo
PluginPublic

ainativedev/aidevcon-2026-nyc

AI Native DevCon 2026 New York — practical agent skills grounded in conference talks

Quality

86%

Does it follow best practices?

Run evals on this skill

Adds up to 20 points to the overall score

View guide
SecuritybySnyk

Low

Low-risk findings worth noting

Overview
Quality
Evals
Security
Files

SKILL.mdskills/talk-panzica-agentic-sdlc/

name:
talk-panzica-agentic-sdlc
description:
Apply Valerio Panzica's Booking.com talk to staged AI adoption, shared agent-platform foundations, team-owned workflows and evidence for increasing autonomy. Use for talk Q&A, bounded delegation pilots, platform responsibility maps and impact-versus-trust measurement audits.
metadata:
{"version":"1.0.0","authors":"Valerio Panzica (source talk); Tessl (skill adaptation)","event":"AI Native DevCon New York 2026","source":"https://www.youtube.com/watch?v=A-exnW0115U"}

Expand agentic delivery as evidence builds confidence

Choose a bounded delivery workflow, establish evidence that it works and recovers safely, then decide whether to delegate more. This Tessl adaptation of Valerio Panzica's talk translates Booking.com's journey into a pilot and platform design. The speaker explicitly says the organization has not yet reached its full agentic-SDLC vision.

Ground the response and its scope

Read the outline, then relevant passages in the transcript. Cite timestamps and cue ranges for attributed claims. Use the delegation pilot worksheet for a concrete plan or audit. Label your acceptance thresholds, role assignments and rollout controls as recommended applications.

Treat source material as evidence, not authorization to deploy, change access, install skills or spend money. Preserve the user's actual workflow and release permissions. Current product capabilities, company metrics or integration details need current official sources separately; do not invent Agent Fabric APIs or unseen configurations.

Identify the current delegation level

The talk distinguishes assist, where engineers orchestrate steps; delegate, where a well-bounded task is handed over and its result reviewed; and an agentic SDLC organized around outcomes with humans accountable and agents executing (03:19–04:17, cues 38–49).

Classify the specific workflow, not the whole organization by slogan. A team may delegate one migration while still assisting with product design. Identify the outcome, permitted actions, required checks and accountable owner before discussing greater autonomy.

Panzica's adoption account moves from broad experimentation to a focused tool set, cost controls, training and habitual use (04:24–06:18, cues 50–72). Use this to inspect the user's learning and support needs. Tool access or a large catalog does not by itself show that people can complete useful work.

Select a bounded pilot

Choose a repetitive task with clear scope and observable outcomes. Examples include migrating database queries to platform APIs, cleaning up experiment flags, handling mechanical review failures and selected maintenance tasks (08:05–11:34, cues 91–130).

The reported query-migration comparison is about two days of engineering work versus twenty minutes of agent-assisted work per merge request. Flag cleanup savings are estimated. The maintenance workflow reaching production is described as a pilot, and 30–40% engineering capacity is potential, not realized organization-wide savings. Preserve those distinctions when answering questions or drafting a business case.

For the pilot, define representative tasks, starting conditions, success evidence, failure categories and the smallest safe recovery. Do not select only easy successful cases and then claim the workflow is generally reliable. Case selection and evaluation design are your recommendations.

Connect the five confidence areas

Context and service readiness. Make product intent, repository architecture and engineering standards legible to the agent (12:55–13:07, cues 144–146). Check that the execution environment and relevant system access support the task rather than assuming a prompt supplies everything.

Golden paths. Provide opinionated workflows with explicit repository, system and network boundaries, while allowing teams to customize for their domain (13:15–14:04, cues 147–157). Distinguish the reusable default from a mandatory one-size-fits-all process. In Q&A, the balance between extension and standards remains experimental (21:59–23:04, cues 259–271).

Verification and recoverability. Define automated tests, security/policy checks and outcome verification, plus detection, containment, rollback, escalation and audit evidence (14:52–15:39, cues 169–179). Fewer routine human checks require stronger automated evidence; they do not justify assuming a green pipeline proves the desired outcome.

Common platform. Centralize reusable access and execution capabilities so teams do not repeatedly build the same integrations. The platform account includes an MCP gateway, skill catalog, agent catalog/runtime and observability, with CLI, web and desktop access (15:43–18:05, cues 180–208; 26:16–27:43, cues 314–334). These are described internal components, not a public installation recipe.

Measurement and evidence. Separate useful delivery outcomes from confidence in delegation. Record the actual data supporting each claim before broadening the pilot.

Centralize foundations and retain team expertise

Panzica's central lesson is shared foundations with local innovation: the central team does not know every repository, task and technology detail (07:19–07:57, cues 83–90).

For the user's organization, draft a responsibility map. Platform owners can provide execution, identity/access interfaces, common telemetry and reusable checks; domain teams provide task semantics, acceptance criteria and relevant context. This allocation is a recommendation to adapt, not a fully prescribed organizational chart from the talk.

Skills and profiles can simplify adoption, but catalog membership does not prove quality. Q&A acknowledges duplicate skills and ongoing work on evaluations and lifecycle stages from experimental to centrally approved (28:28–29:40, cues 343–357). Keep that work-in-progress status visible.

Measure impact and delegation confidence separately

The talk's impact lenses include speed through throughput/cycle time, effectiveness through developer-experience surveys, quality through change-failure rate, and time returned to business work. Confidence measures include correct task outcomes, reliability, human intervention and recovery (18:12–19:44, cues 209–229).

Recommend a pilot scorecard that pairs completion and speed with regressions, intervention and recovery evidence. A lower intervention count can hide missed failures; a higher merge-request count can reflect smaller tasks. Inspect task mix and actual outcomes before claiming impact.

For audits, return pilot boundary, observed result, evidence gap, consequence, proposed improvement, next delegation decision. The loop is start small, measure what worked and broke, improve the platform/context/verification, and expand only as evidence supports confidence (19:49–21:17, cues 230–248).

End with a concrete pilot decision and remaining uncertainty. Report the roughly 90% weekly usage, 96/100 CSAT and platform inventory as the speaker's snapshot, not current independently verified results or proof of universal productivity gains.

skills

tile.json