FIELD NOTE / LINKEDIN
Start routing with the job.
The complete written thought and the evidence behind it. The video edition will follow its public release.
The written argument is here.
This approved LinkedIn edition is on the journal now. Its video player and original platform link will appear after each public release is verified.
Start routing with the job.
Video caption
Start routing with the job.
A fast route can be wrong when none of its choices fits.
Include exit routes and end-to-end outcome checks.
My rule: The fastest eligible system must pass the same contract.
#EricFieldNotes
Full written post / accessibility read
Code, Jev, Haiku and a small Qwen solve different pieces of an agent system. Start by naming the decision class, required output, data boundary, cost of error and latency budget. Then decide which component may attempt it.
Consider an entitlement workflow with an unfamiliar partner exception. A typed classifier must choose from the provided options unless you designed an outside-scope path. A generative model may explain uncertainty, but it still cannot grant entitlement authority.
For each class, specify the allowed route, required state fields, abstention rule, tool authority and independent outcome. Replay held-out routine and unknown cases through each candidate. Measure accepted decisions and rescue time, not just response format.
Use deterministic code where the rule is fixed. Use a decision model for bounded semantics only if it passes the domain test. Use a writer for explanations and an owner for novel exceptions. Do this because routing is an authority decision before it is an optimization.
#EricFieldNotes
Four-beat scene transcript
1. Start routing with the job.
Code, Jev, Haiku and a small Qwen solve different pieces of an agent system. Start by naming the decision class, required output, data boundary, cost of error and latency budget. Then decide which component may attempt it.
Visual: One leaderboard cannot assign business authority.
2. The unknown case breaks a neat router.
Consider an entitlement workflow with an unfamiliar partner exception. A typed classifier must choose from the provided options unless you designed an outside-scope path. A generative model may explain uncertainty, but it still cannot grant entitlement authority.
Visual: A fast route can be wrong when none of its choices fits.
3. Make the table executable.
For each class, specify the allowed route, required state fields, abstention rule, tool authority and independent outcome. Replay held-out routine and unknown cases through each candidate. Measure accepted decisions and rescue time, not just response format.
Visual: Include exit routes and end-to-end outcome checks.
4. Route by cost of being wrong.
Use deterministic code where the rule is fixed. Use a decision model for bounded semantics only if it passes the domain test. Use a writer for explanations and an owner for novel exceptions. Do this because routing is an authority decision before it is an optimization.
Visual: The fastest eligible system must pass the same contract.
Research and claim limits
- TypeSafe AI: Jev documentation (S98)
- TypeSafe AI: Jev 1.13 jaggedness (S99)
- Anthropic: Models overview (S104)
- Qwen: Qwen3 model family (S105)
The examples identified as illustrative or simulated are design probes, not reported incidents. Vendor specifications do not establish workload performance.