Add Idea: Agent Failure Checkpointing (powered by Platform)

ideas/agent-failure-checkpointing.md

@@ -0,0 +1,83 @@
+---
+title: Agent Failure Checkpointing
+author: Arsalan Khan (@asalan316)
+date: 2026-08-13
+tags: [runtime-lifecycle, sandboxing-isolation]
+---
+
+## The idea
+
+CF supports automatic checkpointing of agent state at configurable points, enabling resume-from-failure for long-running agent tasks. Just as CF restarts crashed apps transparently, it should resume crashed agents from their last checkpoint.
+
+## Why it might matter
+
+LangGraph, Anthropic, and multiple frameworks emphasize persistence through failures as critical for production agents. Agent workloads are often multi-step (research, analysis, code generation, testing) and expensive to restart — both in compute cost and wall-clock time. A 30-minute agent task that fails at minute 25 shouldn't restart from scratch.
+
+Platform-level checkpointing reduces application complexity (no custom state management), enables consistent recovery semantics across frameworks, and could integrate with the graduated lifecycle states (running → suspended → dehydrated) for cost optimization.
+
+## Example: Invoice processing agent
+
+A finance team deploys an autonomous agent app to CF that processes vendor invoices overnight:
+
+```text
+cf push invoice-agent -m 1G --health-check-type process
+cf bind-service invoice-agent llm-provider
+cf bind-service invoice-agent s3-documents
+cf set-env invoice-agent CHECKPOINT_STRATEGY "per-task"
+```
+
+The agent processes 500 invoices autonomously. At invoice #387, LLM provider has outage:
+
+```text
+Invoice Processing Agent (deployed app, running autonomously)
+├── Task: Process 500 pending invoices from S3 bucket
+│
+├── Invoice #1-386: extracted, validated, posted to ERP
+│   ✓ checkpoint: {processed: 386, errors: 12, last_id: "INV-2026-0386"}
+│
+├── Invoice #387: fetch from S3, call LLM for extraction
+│   ✗ CRASH — LLM provider 503 at 3:47am
+│
+│   --- CF detects crash, finds checkpoint ---
+│   --- Restarts agent container, restores state ---
+│
+├── Resume from checkpoint (not invoice #1):
+│   Invoice #387: retry extraction
+│   ✓ checkpoint: {processed: 387, ...}
+│
+├── Invoice #388-500: continue processing
+│   ✓ DONE — 4.2 hours total (not 8+ hours if restarted)
+│
+└── Result: 500 invoices processed, 18 flagged for review
+```
+
+**Without platform checkpointing**: Developer writes custom state persistence to Redis/Postgres, handles resume logic, manages partial failure states manually.
+
+**With platform checkpointing**: Developer declares checkpoint strategy in manifest, CF handles persistence and resume transparently.
+
+```yaml
+# manifest.yml
+applications:
+- name: invoice-agent
+  agent:
+    checkpoint:
+      strategy: per-task      # per-task | per-tool-call | explicit | time-based
+      storage: platform       # platform-managed or bring-your-own
+      retention: 7d
+```
+
+**Checkpoint contents** (platform-managed):
+- Agent memory/conversation state
+- Accumulated task outputs
+- Position in work queue
+- Bound service session state (where possible)
+
+## What to research next
+
+- What is the right checkpoint granularity — per tool call, per LLM response, configurable by application?
+- How does checkpoint storage interact with multi-tenancy — per-org storage quotas?
+- What state needs checkpointing — conversation history, tool outputs, environment state, or all three?
+
+## Related
+
+- [per-session-sandboxes](per-session-sandboxes.md) — graduated lifecycle states (running, suspended, dehydrated)