{"resourceId":"nvidia-doca-agent-skills-20261001","versions":[{"version":"external-537dd86d75ce2f193df6ba545d0045c5eb1ba8674c33daf323c6a779f837b88b","resource":{"id":"nvidia-doca-agent-skills-20261001","title":"NVIDIA: DOCA AI agent skills claim 19%→100% checklist pass on 65 BlueField prompts","organization":"NVIDIA","sector":"DPU / infrastructure developer tools","geography":"Global","publishedAt":"2026-10-01T00:00:00Z","publicationDate":"2026-10-01","eventDate":"2026-10-01","sourceName":"NVIDIA Technical Blog","sourceLabel":"Vendor technical blog with internal evaluation","sourceUrl":"https://developer.nvidia.com/blog/build-applications-on-nvidia-bluefield-faster-with-nvidia-doca-agent-skills/","evidenceClass":"vendor-claim","outcomeClass":"emerging","topics":["developers-agents","infrastructure","data-security","operating-model"],"finding":"Oct 1, 2026 NVIDIA Technical Blog announces DOCA AI agent skills on GitHub (SKILL.md packs with verified API signatures, hardware capability requirements, and build constraints spanning Flow, GPUNetIO, PCC, RDMA). NVIDIA reports a 65-prompt internal evaluation where agents without skills satisfied only 19% of graded checklist items versus 100% with skills; common failures without skills included API misuse (59/65), unverified hardware capability (46/65), and wrong tool routing (39/65). A side-by-side RDMA-on-BlueField-3 demo claimed 73% less handwritten code (189 vs 695 lines) and 46% fewer hardware commands (20 vs 37). Skills also emphasize preflight/rollback for mlxconfig-class firmware changes.","sledRelevance":"Relevant to SLED teams building or hardening BlueField-based agent infrastructure (including Open Agent Safety/Sentry paths): shows domain-skill scaffolding can reduce agent coding errors, but evidence is NVIDIA-run, not an independent production study.","evidence":"Opened NVIDIA developer blog dated Oct 01, 2026. Body states 65-prompt evaluation with 19% vs 100% checklist results, failure-mode counts, and 73%/46% RDMA demo reductions; GitHub NVIDIA/skills referenced.","architectureImplications":"Agent-assisted DOCA development should load version-pinned skill packs and enforce hardware capability checks before code generation; firmware-class changes need cold power-cycle awareness.","governanceImplications":"Vendor-owned skill definitions create a currency and trust dependency—require version pinning, review, and change control when agents touch production DPUs.","securityPrivacyImplications":"Skills that encode safe firmware procedures can reduce risky agent actions on live BlueField, but do not replace human approval for production changes.","caveats":"NVIDIA evaluates its own skills; checklist methodology not independently audited; BlueField-3 demo is illustrative; production SLED effectiveness not shown.","streamIds":["nvidia"],"roles":{"sales":"Open with a measured-but-vendor-run developer productivity claim: NVIDIA’s Oct 1 DOCA agent skills blog says agents hit 100% of graded checklist items across 65 prompts with skills versus 19% without, plus a BlueField-3 RDMA demo with 73% less handwritten code. Useful for platform engineering and DPU security teams adopting coding agents on BlueField. Discovery: are agents already writing DOCA/Flow/RDMA code, and who owns skill-pack version control? Unsupported: independent replication, guaranteed production defect reduction, or SLED deploy outcomes.","engineering":"Fit when AI coding agents touch DOCA on BlueField for networking, GPUNetIO, or security services. Integrate SKILL.md packs from NVIDIA/skills with pinned DOCA versions and mandatory capability preflights. Prerequisites: BlueField lab, DOCA toolchain, agent harness that loads skills. Constraints: vendor-owned eval; skills must track DOCA releases; firmware writes still need human change control. Proof: rerun a local prompt suite graded with your checklist before trusting 19%→100%.","delivery":"Work: stand up a BlueField agent-coding sandbox with pinned DOCA skills, human approval for mlxconfig-class changes, and a regression prompt suite. Dependencies: DPU platform team, InfoSec, developer experience. Ownership: infrastructure platform engineering. Skills: DOCA + agent harness ops. Governance: skill version review each DOCA upgrade. Acceptance: local checklist pass-rate target and zero unapproved firmware writes. Risks: agents inventing APIs if skills drift; over-trusting vendor demo line counts."},"roleDetails":{"sales":{"customerProblem":"DPU/platform teams want safer AI coding on BlueField without believing marketing demos wholesale.","stakeholders":["infrastructure platform","network engineering","CISO","developer experience"],"discoveryQuestions":["Are agents writing DOCA today?","Who pins skill versions?","Lab BlueField available for local eval?"],"valueHypothesis":"Vendor 65-prompt results can justify a bounded skills pilot if locally re-graded.","potentialEngagement":"BlueField agent-coding pilot scoping with checklist design workshop.","unsupportedClaims":["Independent 100% claim verification","Production incident reduction guarantee","ATO from this blog"]},"engineering":{"fit":"Fit for agent-assisted DOCA development on BlueField-3/4 fleets.","architectureIntegration":"Load NVIDIA/skills DOCA packs into the coding agent; enforce capability checks and pkg-config/build constraints before apply.","prerequisites":["DOCA version pin","BlueField lab","agent harness with skill loading","firmware change window process"],"constraints":["Vendor-only evaluation","skill currency risk","not a production SLED study"],"security":"Firmware-oriented skills reduce but do not eliminate need for OOB paths and human approval.","proofOfValue":["Replicate prompt suite locally","Track API-invention rate","Require cold-cycle awareness in runbooks"]},"delivery":{"work":"Deliver a controlled agent-coding sandbox for DOCA with change-management for hardware-affecting steps.","dependencies":["DPU platform","DevEx","InfoSec","release management"],"ownership":"Infrastructure platform owns skills lifecycle; security owns production-change gates.","skillsAdoption":"Train developers to verify agent output against DOCA docs/skills, not trust first compile.","governanceCheckpoints":["Skill version approval","Firmware change CAB","Prompt-suite regression on DOCA upgrade"],"acceptanceCriteria":["Local checklist pass-rate report","No unapproved mlxconfig writes","Documented rollback drill"],"risks":["Skills drift after DOCA upgrade","Agents applying firmware changes outside maintenance windows"]}},"retrievedAt":"2026-10-02T03:05:47Z","enrichedAt":"2026-10-02T03:09:13Z","enrichmentBasis":"retrieved source","accessibilityWorkforceImplications":"Skill-augmented agents can widen who can attempt DOCA work; still require accessible review checklists so non-experts do not approve unsafe firmware steps.","procurementImplications":"If buying BlueField/DOCA support, ask vendors how agent skills are versioned, tested, and updated across DOCA releases.","operatingModelImplications":"Treat agent skills as operational artifacts under the same change control as infrastructure-as-code.","sourceVerification":{"openedUrl":"https://developer.nvidia.com/blog/build-applications-on-nvidia-bluefield-faster-with-nvidia-doca-agent-skills/","referenceExcerpt":"Without skills, agents satisfied only 19% of graded checklist items on real DOCA tasks.","promptVersion":"sled-research-v4.0","model":null,"basis":"agent-reported inspection"}}}]}