Continual-learning agents

Agents that get better at your work by doing it, without waiting for a lab to retrain them.

Facts

Lane: Agents and organisations
Wave: 8 of 12 (build order, not a date)
Needs: 20 boxes all the way back through hard links
Unlocks: 2 boxes downstream through hard links
Leverage: 0.2 points · stands in the way of 2 of 53 stages · rank 38 of 57
Cluster: Agents at work
When: Sourced forecast: 2026 to 2035

Evidence as of October 2026

Dwarkesh Patel and Karpathy both name continual learning as the main missing piece. Silver and Sutton argue the next era learns from experience rather than human data.

Watch
Whether any lab ships weight-level learning from deployment.

Timeline note
Karpathy said in October 2025 that agents need about a decade, with continual learning a main missing piece.

01Andrej Karpathy on the Dwarkesh Podcast

Around it

Needs · 7

Knowledge-rich agentsstage · Emerging
Trace capturecapability · Emerging
Reliable verifierscapability · Built today
Correction propagationcapability · Emerging
Weight-level continual learningcapability · Emerging · helps, not required
Robust generalisationcapability · Emerging · helps, not required
Firm-owned post-trained modelscapability · Emerging · helps, not required

Unlocks · 4

Drop-in remote workerstage · Speculative
Software that maintains itselfstage · Speculative · helps, not required
Agent-native organisationsstage · Emerging · helps, not required
Agent-run discoverystage · Emerging · helps, not required

Loops and off-ramps

Learning from deploymentreinforcing loop
Capex pullbackoff-ramp · cut off
Taiwan chip disruptionoff-ramp · cut off
Power shortfalloff-ramp · cut off
Capability plateauoff-ramp · cut off
Training data ruled outoff-ramp · slowed
Misuse catastropheoff-ramp · slowed

Openings researchers' judgement, October 2026 scouting

Startup room 37/100 · rank 79 of 124 Research room 61/100 · rank 26 of 124

Crowding3 Lab overlap4 Small-team fit3 Buyer today2.5 Research openness5

Depth: light
Players: Trajectory ($40m at $300m, Sequoia); NeoCognition ($40m seed)
Wedge: Safe online weight updates from production data without forgetting or poisoning (see contlearn). research

01AI Weekly: Trajectory $40m