Multi-LLM orchestration without the drift.

Your multi-agent pilot works in demos and breaks in production, and it is almost never the model's fault. It is the handovers. CALEL Legion is the framework we use to build multi-LLM systems where the handovers hold.

1. Handover drift is a measurable failure.

The MAST taxonomy published at NeurIPS 2025 analysed 1,600+ multi-agent execution traces and found three root-cause categories, each accounting for roughly a third of failures: specification ambiguity, coordination breakdowns, and verification gaps. All three converge at the same physical place in the system — the seam where one LLM hands work to another. Augment Code's summary of the taxonomy is worth reading in full.

Anthropic's own guidance for long-running agents names the same two failure modes explicitly: context-loss-induced incoherence and premature wrap-up near context limits. Their recommendation is a structured handoff to a fresh agent, not a longer window. The fix is not more tokens. It is a better seam.

2. One orchestrator, many models, one context.

CALEL Legion treats the orchestrator as a first-class object with a persistent, structured context that outlives any individual model call. Claude for reasoning, GPT for tool-use loops, Gemini for speed, a distilled Llama for the cheap paths — each is a worker. The orchestrator keeps the goal, the constraints, the running artefact and the failure history in one place, and hands each worker only what it needs.

The pattern reads a bit like an intelligent router with memory. RouteLLM and its cousins already deliver up to 85% cost reduction whilst holding 95% of GPT-4 performance by matching workloads to models. CALEL Legion adds the missing piece: the context that persists across the routing decisions, so the third worker knows what the first one decided.

3. The eval loop applies to agents too.

The same discipline we use for Voice AI — a small, hand-labelled eval suite wired in on day two — applies to agent systems. Twenty end-to-end task traces. Every commit runs against them. If pass rate drops, the commit does not merge. That is how we catch the silent quality degradation that Gartner is now describing as the reason 40% of agentic pilots will be cancelled by the end of 2027.

If your agent achieves 85% per-action accuracy, a ten-step workflow succeeds about 20% of the time. The eval loop is not optional.

4. MCP wiring is a plumbing problem, not a design problem.

The Model Context Protocol has already collapsed the N×M integration problem to N+M — every model can call every tool through one JSON-RPC contract instead of a matrix of custom shims. That is a real gain. It is also a plumbing gain, not a design one; the hard question is still which tool to call, in what order, with what verification. CALEL Legion assumes MCP as the transport and spends its complexity on the routing and verification layer above it.

• Persistent orchestrator context across model handovers
• Structured task specifications, so workers cannot misinterpret their role
• An eval harness wired in from day two, not bolted on before launch
• MCP as the tool-call transport, not something you re-implement per model
• Loud failure at the seam — the orchestrator refuses to proceed on a bad handover instead of quietly cascading

5. Fail loudly at the seam.

The single most valuable thing an orchestrator does is refuse to accept a bad handover. If Worker B returns something that does not match the spec Worker A produced, the orchestrator stops, logs, escalates. It does not paper over the inconsistency and hand it to Worker C. Coordination breakdowns and verification gaps — two-thirds of the MAST failure modes — are what you get when the orchestrator quietly forgives at the seam.

The trade.

You give up the "single giant prompt in LangChain" approach that felt fast at demo stage. In exchange, you get a system where you can swap models, add tools, and know at any commit whether the whole thing still works — because the eval loop tells you, and the orchestrator refuses to hide the answer.

Sources.

1. Augment Code — Multi-Agent AI Systems: Why They Fail and How to Fix Coordination Issues (2026).
2. Drew Breunig — Enterprise Agents Have a Reliability Problem (2025).
3. Swfte — Intelligent LLM Routing: How Multi-Model AI Cuts Costs by 85% (2026).
4. SitePoint — MCP: Complete Guide for AI Integration (2026).
5. Hacker News — Ask HN: Is it just me or LangChain/LangGraph DevEx horrible? (2024, still the reference thread).