ANTHROPIC DEBUTS MANAGED AGENTS AND SHIPS CLAUDE CODE 2.1.108/109 WITH PROMPT CACHING CONTROLS AND SESSION RECAP
Anthropic introduced Managed Agents with stable agent interfaces and updated Claude Code with prompt caching controls and a session recap feature. Anthropic’s ...
Anthropic introduced Managed Agents with stable agent interfaces and updated Claude Code with prompt caching controls and a session recap feature.
Anthropic’s engineering team detailed Managed Agents, a hosted service that virtualizes agent components—session, harness, and sandbox—so implementations can change without breaking user code. The post argues this decoupling keeps agent workflows stable as models evolve, avoiding “pet” containers and stale harness assumptions blog.
In parallel, Claude Code v2.1.108 adds environment flags for prompt caching TTLs across API key, Bedrock, Vertex, and Foundry, a /recap feature to regain context after breaks, skill-based discovery of built‑in commands, better error surfacing, and memory-use cuts. v2.1.109 is a minor UI tweak for the extended‑thinking indicator (2.1.108, releases).
Stable agent interfaces mean fewer rewrites when models or toolchains change, reducing maintenance drag on long‑horizon automations.
Prompt caching controls and session recap can lower cost and latency while keeping multi‑day engineering sessions coherent.
-
terminal
Enable ENABLE_PROMPT_CACHING_1H vs FORCE_PROMPT_CACHING_5M and measure cost, latency, and cache hit rate across Bedrock, Vertex, and Foundry-backed runs.
-
terminal
Run a long task under Managed Agents, then swap or update the harness; verify session continuity, sandbox isolation, and tool-call routing stay correct.
Legacy codebase integration strategies...
- 01.
If you have homegrown agents, evaluate carving out session/harness/sandbox as separate services to de-risk model upgrades; avoid single ‘pet’ containers.
- 02.
Update ops to new caching env vars; deprecations like ENABLE_PROMPT_CACHING_1H_BEDROCK are still honored but should be migrated.
Fresh architecture paradigms...
- 01.
Design agent workflows around session logs and sandboxed execution from day one, so you can swap harnesses or models without rewrites.
- 02.
Use skill-discovered slash commands (/init, /review, /security-review) and /recap to script repeatable routines and resumable work.