Amazon S3 Annotations: 1GB Metadata per Object for AI Agent Workflows

AWS today announced Amazon S3 annotations, a new metadata capability that lets you attach up to 1,000 named annotations per object, each up to 1 MB in size, totaling up to 1 GB per object. Annotations support JSON, XML, YAML, or plain text, and can be modified or deleted without rewriting the object. The feature is designed for AI agent workflows that need rich, evolving context at scale.
Key Details
- Scale: 1,000 annotations per object, each 1 MB max, total 1 GB per object.
- Formats: JSON, XML, YAML, plain text.
- Mutability: Add, modify, or delete annotations at any time without touching the object data.
- Lifecycle: Annotations move with the object during copy, replication, and cross-region transfers; removed on object deletion.
- Queryability: When S3 Metadata is enabled, annotations flow into annotation tables queryable via Amazon Athena and other analytics engines. AI agents can discover data using natural language through the S3 Tables MCP server.
- Storage classes: Query annotations on objects in any storage class, including Glacier, without restoring objects or paying retrieval charges.
Using the CLI
Create a JSON annotation on an existing object:
# Create a JSON file with technical metadata
cat > mediainfo.json << 'EOF'
{"codec":"H.265","resolution":"3840x2160","audio_tracks":8,"frame_rate":29.97}
EOF
Attach it as an annotation
aws s3api put-object-annotation
--bucket my-media-bucket
--key videos/documentary-2026.mp4
--annotation-name mediainfo
--annotation-payload ./mediainfo.json
Attach a plain-text AI-generated summary as a separate annotation:
echo "A 90-minute nature documentary..." > summary.txt aws s3api put-object-annotation
--bucket my-media-bucket
--key videos/documentary-2026.mp4
--annotation-name ai-summary
--annotation-payload ./summary.txt
Use Cases
- Media & Entertainment: Track transcripts, content moderation results, subtitles, and licensing metadata as separate annotations on video assets.
- Financial Services: Attach AI-generated investment summaries to research documents for autonomous research agents.
- Life Sciences: Annotate clinical trial data with regulatory status, patient cohort details, and approval chains — accessible even in Glacier without retrieval charges.
Permissions
IAM policies need s3:PutObjectAnnotation and s3:GetObjectAnnotation actions. No changes to the object itself are required.
Compared to Existing Metadata
| Capability | Max Size | Mutable? | Best for |
|---|---|---|---|
| System-defined metadata | Fixed | No | Object properties (size, storage class, creation time) |
| User-defined metadata | 2 KB | No (set at upload) | Small custom key-value pairs |
| Object tags | 10 tags, 128/256 chars per key/value | Yes | Access control, lifecycle rules, cost allocation |
| Annotations | 1 GB (1,000 × 1 MB) | Yes | Rich business context (JSON, XML, YAML, plain text) |
📖 Read the full source: HN AI Agents
👀 See Also

Qwen 3.6 27B Benchmarked on DeepSWE: 2% Score, 70 Hours, 44k Avg Output Tokens
Qwen 3.6 27B (FP8, BF16 KV cache, 262k context) scored 2% on DeepSWE in 70 hours. Output tokens averaged 44k per task — comparable to larger models like Qwen 3.6 Plus. Ran on 1x RTX6000 Pro Blackwell via RunPod.

Analysis: Anthropic's actual compute costs for Claude Code users are far lower than reported $5k figure
A recent article analyzes the claim that Anthropic's $200/month Claude Code Max plan consumes $5,000 in compute, finding that actual inference costs are roughly 10% of API prices when comparing to competitive open-weight models on OpenRouter.

OpenAI Working on AI Smartphone with MediaTek/Qualcomm Chips; Mass Production Target 2028
According to supply chain analyst Ming-Chi Kuo, OpenAI is developing an AI smartphone with chip partners MediaTek and Qualcomm, exclusive manufacturer Luxshare Precision, and mass production planned for 2028. The device is positioned as a context-aware AI agent platform.

AI Coding Agents Struggle with Context Management in Large Codebases
Analysis of AI coding agents reveals they spend 15-20 tool calls on orientation tasks like grepping for routes and reading middleware before writing code, burning through context windows. Vercel achieved 100% accuracy by stripping 80% of tools and using bash, while Pi uses just 4 tools and a system prompt under 1,000 tokens.