Building Agents with Claude Opus 4.8: Live Demos with Letta and Browserbase

Opus 4.8 introduces a five-level effort system that replaces the binary "think hard / think fast" of earlier models:

{"effort": "low" | "medium" | "high" | "xhigh" | "max"}

• high is the new default (was medium). Same token budget as before, better results. This is the "free upgrade" tier.
• medium costs meaningfully less. For routine tasks (the daily audit, style checks) this is the right choice.
• xhigh and max are new. max enables Ultracode: dynamic multi-agent workflows for latency-insensitive tasks.

The cost implications for our LiteLLM proxy routing: the daily audit should use medium (cheapest adequate model), brief generation uses high (default), and formal verification tasks use xhigh or max.

4.8 reports progress between tool calls and writes fuller wrap-ups without scaffolding. The old pattern of "summarize every N calls" is no longer needed. A one-line silence default recovers 4.7-style quiet output with no quality change.

This connects to the REPL gossip model¹: if the agent narrates its progress, participants sharing the REPL heap see intermediate state without polling.

The new top of the effort menu: xhigh effort + dynamic multi-agent workflows. Claude Code spawns subagents for the hardest, latency-insensitive tasks. This is the mechanism behind the "Running 3 agents" display in the Letta demo.

The key architecture from Charles Packer's demo. Letta (formerly MemGPT²) provides stateful agents with persistent memory via a filesystem abstraction:

LLMs have limited context, so use LLMs to manage their own context. Memory subagents rewrite context over time for continual learning in token space. Structure memory in clear hierarchies:

system/
├── human/
│   ├── prefs/
│   │   ├── coding_style.md
│   │   ├── communication.md
│   │   └── workflow.md
│   └── identity.md
├── project/
│   ├── gotchas.md
│   └── overview.md
├── persona.md
├── history/
│   └── project_timeline.md
├── claude.md
├── codex.md
└── api_documentation.md

Our equivalent: CLAUDE.md + .claude/ + .beads/ + auto-memory files in ~/.claude/projects/. The hierarchy is similar; the format differs (Letta uses FS paths, we use org property drawers + JSONL).

"The most powerful way for an agent to interact with anything is via bash… including interacting with memory/context."

• Programmatic modification of memory/context
• Batch operations to restructure memory
• Coding is "in distribution"³

"How can we prevent agents from being 'stuck' on a local machine?"

From 1 agent : 1 machine to 1 agent : N machines.

This is exactly the problem the shared REPL solves: the nREPL on nexus is reachable from laptop, hydra, and any LAN machine. The agent's state (loaded namespaces, bound vars) is on the server, not the client.

Letta Code agents live on the server (state/memory), but can move between deployment environments:

• Macbook → Letta Code server
• GitHub Action → Letta Code server
• Cloud Sandbox → Letta Code server

Server stores: Agent state (Postgres) + Context Repo (git + GCS). Each agent has a UUID: agent-0c771d18-bc8f-4ae2-81a1-f64508688db1.

System-prompt editing is key to continual learning (in token space):

System Prompt → LLM → Experience
      ↑                    │
      └── Update memory ───┘
         based on new experiences

This is the feedback loop our daily-publish audit implements: audit brief → annotate with verdict → next audit reads the annotation.

Cache hits are 10x cheaper than base input. The Letta memory approach benefits from this: the system prompt (with memory) is cached; only the new experience invalidates the suffix.

New in 4.8: {"role": "system"} messages after a user turn. Mid-conversation system messages have the same authority as the top-level system field but don't invalidate the cached prefix. Use for instructions that become relevant later.

Ref: platform.claude.com/docs/…/mid-conversation-system-messages

The Letta UI shows memory as nodes (PERSONA, ONBOARDING, HUMAN) with recent memory changes tracked:

• "Dreaming" state: chore(reflection): update onboarding tracking
• "Tutor" role: Save the user's name as provided
• Git-style commits: "Initial commit" per user session

The "Reflected on /palace, the halls remember more now" message suggests a method-of-loci metaphor for memory organization.

Browserbase provides managed headless Chrome infrastructure. Agents get a browser session via API, execute actions (click, type, navigate, screenshot), and return structured results.