Reliability on repeat

Getting the same task right every time, not once.

Facts

Lane: Models and thresholds
Wave: 6 of 12 (build order, not a date)
Needs: 0 boxes all the way back through hard links
Unlocks: 14 boxes downstream through hard links
Leverage: 2.3 points · stands in the way of 12 of 53 stages · rank 2 of 57
Cluster: Agents at work

Evidence as of October 2026

On τ-bench, GPT-4o succeeded on over 60% of retail tasks but on under 25% when it had to succeed eight times in a row. Success rates also decay with task length.

Around it

Needs · 1

Calibrated abstentioncapability · Emerging · helps, not required

Unlocks · 5

AGI: top-expert-level AIstage · Speculative
Agent-native organisationsstage · Emerging
Automated coderstage · Emerging
Drop-in remote workerstage · Speculative
Robotaxis at scalestage · Built today · helps, not required

Loops and off-ramps

Trust and delegationreinforcing loop

Openings researchers' judgement, October 2026 scouting

Startup room 100/100 · rank 1 of 124 Research room 100/100 · rank 1 of 124

Crowding4 Lab overlap4 Small-team fit4 Buyer today4 Research openness5

Depth: deep
Players: Braintrust ($80m B, $800m); LangChain ($1.25bn); Patronus ($50m B); LMArena ($1.7bn); Princeton CITP research
Wedge: Runtime verifier and recovery layer that raises effective pass^k on a fixed model for one workflow; research on methods that raise consistency and replay-proof benchmarks. startup