AI Native DevCon 2026 New York — practical agent skills grounded in conference talks
92%
Does it follow best practices?
Run evals on this skill
Adds up to 20 points to the overall score
Passed
No findings from the security scan
This is a Tessl working template applying the talk. Its fields, ownership model, tests, and proposed measurements are recommendations, not a schema specified by the speaker. Use only the sections that help the user's decision, and leave unavailable evidence explicitly unknown.
State the decision or task, caller, delivery phase, and scope of the material actually inspected.
| Information or instruction | Purpose for this task | Codified / ambient / mixed | Why this lifecycle fits | Source and version/date | Owner and access scope |
|---|---|---|---|---|---|
| To inspect | Needed behavior or knowledge | To decide | Reuse and rate of change | Evidence required | To confirm |
Classify the information rather than assuming every wiki page is ambient or every Markdown file is codified. Split mixed documents if stable instructions and changing facts need different maintenance. Keep essential instructions reliably delivered, and retrieve task-specific evidence when needed.
| Context type | Stage | Design decision | Evidence/check | Owner |
|---|---|---|---|---|
| Codified | Generate | How instructions are authored and reviewed | Versioned source; intended behavior | To assign |
| Codified | Evaluate | Representative tasks and edge cases | Evidence of correct application | To assign |
| Codified | Distribute | How the intended version reaches agents | Installed/delivered version | To assign |
| Codified | Observe | How usage and failures become updates | Traces or user feedback | To assign |
| Ambient | Ingest | Sources, synchronization, backfill, freshness, identity, permissions | Missed-update and access-boundary checks | To assign |
| Ambient | Evaluate | Questions, gold evidence and answers, edge cases | Retrieval coverage and reasoning quality | To assign |
| Ambient | Assemble | Caller/task, relevant sources, conflict handling, budget | Selected evidence, unresolved conflicts, latency/cost | To assign |
| Ambient | Observe | Input quality and engine quality | Disputed sources, poor-answer clusters, regression cases | To assign |
Use code to inspect current behavior, not to automatically overrule a requirement. Do not suppress a real conflict because the answer would be more convenient without it. Retain access boundaries when returning evidence to the caller.
For each case record the question, caller/access scope, expected source documents, expected answer or abstention, and observed result. These cases extend the talk's suggestions into a practical starting set:
Choose project-specific thresholds before interpreting results. The talk does not supply a universal accuracy target, token budget, or synchronization interval. Record what is unmeasured and distinguish correct retrieval from correct reasoning.
| Finding or hypothesis | Evidence | Consequence | Proposed change | Verification |
|---|---|---|---|---|
| To inspect | Source/trace needed | Effect on the task | Smallest useful change | Case that should improve |
Separate changes to input knowledge from changes to the context engine. For example, repeated conflicts may indicate a stale document, an overly strong code preference, or a retrieval failure; inspect before choosing a fix. Turn each confirmed failure into a regression case and monitor whether the issue recurs after the change.
.tessl-plugin
skills
talk-boeckeler-context-lifecycles
talk-chanezon-gordon-agent-experience-evaluation
talk-meppiel-versioned-agent-contracts