Tag: llm

Moving an agent from claude-sonnet-4-6 or claude-opus-4-7 to claude-sonnet-5-5 or claude-opus-5-5 is not just a model ID swap. Thinking is always on, thinking-off and forced tool_choice now return 400, the effort defaults moved, Sonnet counts about 30% more tokens, and Claude Code ignores your old effortLevel on Opus 5.5. Here is the full checklist with before/after code.

Run the orchestrator on claude-opus-5-5, give implementer and reviewer subagents claude-sonnet-5-5, and send search, triage and summarizing work to claude-haiku-5-5 at medium effort. Here is the role matrix, the cost math per subagent run, Haiku 5.5's 100k-token price cliff, and the Claude Code model-resolution rules that silently override your choice.

A claude -p run on Opus with an API key gives up after three consecutive 529s, about 1.5 seconds of backoff, no matter what CLAUDE_CODE_MAX_RETRIES says. Add --fallback-model, set CLAUDE_CODE_RETRY_WATCHDOG=1 with a max wait, or widen the 529 backoff with the OVERLOADED_RETRY delay variables. Checked against Claude Code 2.1.309.

Claude Code 2.1.289 and earlier resend every screenshot on every turn, so one refused request means every later turn is refused too. Update to 2.1.290+ and resume with claude --continue; the client then trims the oldest images to 510, 255, then 127. On an old build, /compact or /clear. Then stop the screenshots piling up.

Magentic orchestration is Agent Framework's port of AutoGen's Magentic-One: an LLM manager writes a fact sheet and plan (the task ledger), then grades every round with a JSON progress ledger to pick the next agent, detect stalls and replan. Here is how the loop works on Microsoft.Agents.AI.Workflows 1.24.0 and agent-framework-orchestrations 1.3.1, what it costs per round, and the limit defaults that let it run forever.

The Agent Harness is Microsoft Agent Framework's batteries-included runtime: one call (AsHarnessAgent in .NET, create_harness_agent in Python) wraps any chat client in a tool loop with todos, plan/execute modes, file memory, tool approval, compaction and OpenTelemetry. Here is what it actually sends to the model on Microsoft.Agents.AI.Harness 1.23.0, what it costs per call, and how to trim it.

A context window is everything the model can see in one request: system prompt, tool definitions, history, tool results, thinking and the reply. Here is what counts, what happens at the limit on claude-opus-5-5 and friends, and the six techniques agents use to work past it: deferred tools, context editing, compaction, memory, subagents and output filtering.

Tool calling is a protocol, not magic: the model emits a structured request matching a JSON Schema you supplied, your code runs it, and the result goes back as a message. The schema and its descriptions are rendered into the model's prompt and, with strict mode, into a decoding grammar, so they shape behaviour more reliably than any system-prompt instruction. Here is how it works on claude-sonnet-5-5 and OpenAI strict mode, with schema patterns that make invalid calls impossible.

In a workflow, your code decides the next step and the model fills in each step. In an agent, the model decides the next step and your code only executes it. That one question (who owns the control flow) determines cost, latency, testability and failure modes. Here is the distinction as Anthropic, Microsoft Agent Framework and LangGraph define it, with the same task built both ways in Python against claude-haiku-4-5 and claude-sonnet-5-5.

An agent skill is a folder with a SKILL.md file that the agent loads on demand: only its name and description sit in the system prompt, the instructions arrive when a task matches, and bundled scripts run without ever entering context. Here is how that differs from a system prompt in cost, scope, invocation and trust, with the exact limits from the Agent Skills spec and Claude Code.