At API prices, the median researcher by agent usage spends over $600/day on inference. The 90th percentile is over $7,000/day. At the start of the year that median was modest coding-agent use. A rising share now run 4+ agents at once, subagents included.

Mid-August 2026 throughput: 3.1 agent workdays per human workday, on an 8-hour standard. Before June 2026, total agent runtime was still below total human labor. Experiments per active experimenter hit an all-time high in August, on data tracked since Jan 2025.

Their bar for "intern": well-defined research tasks under human direction, some of which would take a skilled researcher a few days. They map usage onto Epoch AI's taxonomy: Decide, Design, Build, Run, Analyze, Communicate. All six grew from Jan to Aug.

OpenAI put out "Research acceleration: The view inside OpenAI" on Sep 6. They say they hit the automated research intern milestone they announced last fall. Next target: an automated AI researcher by March 2028. #AI

Ingest is async. Success = accepted; records land after processing (seconds to minutes per the docs). Read back: ListMemoryRecords, RetrieveMemoryRecords, GetMemoryRecord. Kinesis notifies on creation. Failed? ListMemoryExtractionJobs, then StartMemoryExtractionJob to redrive.

Payloads: conversational (USER/ASSISTANT role + text) or JSON (behavioral events, activity logs, system events). Metadata optional, same extraction pipeline as CreateEvent. Fans out to configured strategies, minus self-managed. actorId + sessionId + namespace = shared context.

Sep 8, 2026: IngestData on bedrock-agentcore. Send content straight to long-term extraction. No short-term event gets written. Long-term memory now stands on its own, with no short-term layer underneath. Live in every Region where AgentCore Memory runs.

Until this week, the only way into AgentCore long-term memory was CreateEvent. You wrote a short-term event you'd never read back, just so extraction had something to chew on. AWS cut that step on Sep 8. #AI

Ingest is async. Success = accepted; records land after processing (seconds to minutes per the docs). Read back: ListMemoryRecords, RetrieveMemoryRecords, GetMemoryRecord. Kinesis notifies on creation. Failed? ListMemoryExtractionJobs, then StartMemoryExtractionJob to redrive.

Payloads: conversational (USER/ASSISTANT role + text) or JSON (behavioral events, activity logs, system events). Metadata optional, same extraction pipeline as CreateEvent. Fans out to configured strategies, minus self-managed. actorId + sessionId + namespace = shared context.

Sep 8, 2026: IngestData on bedrock-agentcore. Send content straight to long-term extraction. No short-term event gets written. Long-term memory now stands on its own, with no short-term layer underneath. Live in every Region where AgentCore Memory runs.