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
Use the sections relevant to the user's decision. This is a Tessl working template derived from the talk, not an official Scope configuration. Experimental controls, reporting fields, and acceptance thresholds below are recommended applications and must be adapted to the project.
Record harness, model/version, OS, repository revision, directory path, SDK/version, documentation snapshot, extensions/diagnostics, network access, ordinary tools, credentials available, and reset procedure. Use placeholders for secret values. Record changes between runs.
| Arm | Added context/integration | What remains identical | Question answered |
|---|---|---|---|
| Baseline | No added intervention under test | Scenario, ordinary tools, model, environment, starting state | What happens already? |
| Skill only | Named and versioned skill | Same fixed profile | Does the skill help? |
| MCP only | Named and versioned server | Same fixed profile | Does the server help? |
| Combined, if useful | Same skill plus same server | Same fixed profile | Is the combination helpful or conflicting? |
Define commands or observable checks using the actual project. Do not invent commands where the build or test interface is unknown.
| Gate | Concrete evidence required | Failure treatment |
|---|---|---|
| Language/build validity | Project builds or passes the appropriate language checks | Mark gate failure |
| Tests | Task-specific automated tests pass | Mark gate failure |
| Runtime behavior | App performs the required action against controlled data/service | Mark gate failure |
| Deployment, if required | Deployment succeeds and deployed behavior is checked | Mark failure, or explicitly out of scope |
Keep criterion scores diagnostic when gates fail; a good judge score does not override a nonworking application.
For each criterion, specify expected behavior, the evidence inspected, and pass/fail or scoring rules. For a database workload, consider SDK correctness, workload-appropriate data modeling, partition strategy, query behavior, and safe handling of credentials. These must follow the actual requirements and current product guidance; they are not universal rules from the case study.
Choose runs per arm, order, budget, stopping rules, and treatment of infrastructure failures before execution. Justify sample size for the decision. Keep raw run IDs, code snapshots, traces, gate outputs, criterion evidence, runtime, token usage, and known cost. Record denominators and missing observations. Any statistical intervals or tests are additions to the talk and should be identified as such.
| Arm | Completed / attempted | Gate passes / evaluated | Criterion outcomes (n/N) | Time distribution | Token distribution | Cost, if known |
|---|---|---|---|---|---|---|
| Baseline | Unmeasured | Unmeasured | Unmeasured | Unmeasured | Unmeasured | Unmeasured |
| Intervention | Unmeasured | Unmeasured | Unmeasured | Unmeasured | Unmeasured | Unmeasured |
For each claimed improvement, state the comparable baseline, numerator and denominator, uncertainty, and scope. Report regressions and failures alongside gains. Predetermine a project-specific adoption threshold; do not treat the talk's example percentages as targets or guaranteed outcomes.
| Observed failure and run ID | Source/trace evidence | Proposed change | Rerun and regression check | Owner |
|---|---|---|---|---|
| To investigate | Required evidence | Smallest useful change | Relevant scenario and adjacent cases | To assign |
Close with the next decision the evidence supports and what it does not yet establish.
.tessl-plugin
skills
talk-boeckeler-context-lifecycles
talk-chanezon-gordon-agent-experience-evaluation
talk-meppiel-versioned-agent-contracts