Discuss your workflowThe roadmap starts with boundaries that must exist before the first pilot and ends with federation across teams, agents, and regulated workloads. Phase is sequencing; tier is the source deck’s MVP-versus-full distinction. They are not the same thing.
Before any agent runs.
Single team, contained blast radius.
Pilot needs one hosted model and one self-hosted fallback behind the same interface, both version-pinned, with a documented switch procedure.
Pilot needs authentication, per-tenant quota and token metering. Semantic routing and caching are optimisations for phase 2.
Pilot can ship one channel with streaming. Approval and clarification surfaces become mandatory the moment agents can write to production systems.
Pilot needs injection screening, PII redaction, output validation and workload identity. Full supply-chain control and inter-agent security are mandatory before production data or write access.
Instrument from the first pilot request. Retrofitting correlation IDs after the fact is the most common and most expensive omission in this whole composition.
Pilot needs a golden set and a regression suite fed by real incidents. Adversarial campaigns arrive with production; simulated users and environments are the phase 3 workstream for rehearsing at scale.
Adopt MCP in the pilot to avoid rewriting integrations later. A2A federation and a shared agent registry are the phase 3 workstream, once more than one team runs agents that must delegate to each other.
Pilot needs a single orchestrator with durable checkpoints and a hard step and budget ceiling. Multi-agent topologies only after a single agent has been measured and found insufficient.
Pilot ships hybrid retrieval with reranking and a hard token budget. Add graph retrieval and iterative refinement once you can measure groundedness.
Real systems, real data, real users.
Keep it lightweight in the pilot: an agent registry, an owner per agent and a written autonomy tier. Promote to full control the moment agents touch production systems or personal data — before any traffic, not after.
Ship read-only tools in the pilot. Every write tool should arrive with an approval path, an idempotency key and a rollback story — no exceptions in phase 2.
Start with the twenty terms that appear in disputed answers, not a full enterprise ontology. Expand only where ambiguity has actually cost something.
Connect three sources properly — contract, lineage, entitlements — before connecting thirty carelessly. MCP data servers shorten integration work substantially, but the governance is the slow part, not the plumbing.
Pilot may use session-scoped memory only. Long-lived cross-session memory should not ship before poisoning controls and erasure both work.
Many agents, many teams, regulated load.
These are monitoring perspectives, not five new maturity scores. Define applicable measures and thresholds for the actual workflow.
Each phase is complete when its controls can produce evidence under failure—not when the happy-path demo runs once.