Your Protocol Specifies the Envelope, Not the Infrastructure
MCP and A2A are message-format agreements, and teams that adopt them believing they have solved agent interconnection have quietly signed up to own discovery, queuing, session persistence, identity, and retry semantics themselves — usually without naming those concerns, so they surface as scattered patches inside agent code rather than as a designed layer.
AngleRead the protocols for what they do not guarantee: enumerate the four missing properties (discovery, queues, persistence, stateful sessions), then show where each one reappears as ad hoc code in a typical agent repo. Argue that the correct response is to name the missing guarantees before choosing any product.
Source video ↗The Human Is the Message Bus
In today's agentic workflows the developer has become the router, discovery service, and context serializer between parallel stateful sessions, and that manual relay — not the agents' intelligence — is where latency, token cost, and error injection concentrate; the first thing worth automating away is the human seam, which is an undeclared API that should be designed.
AngleTreat 'loop engineering' as an anti-pattern with a measurable cost: count the copy-paste events, tokens re-serialized, and wall-clock time lost per multi-session task, then show what the seam looks like when it is an actual interface.
Source video ↗Continuity Is a Layer, Not a Variable
Agent process lifetime and conversation lifetime are different things, and systems that keep session state in-process have silently fused them; naming continuity (persistence, hydration, runtime binding) as a first-class layer forces the persistence boundary to be explicit and turns agent runtimes into disposable, replaceable workers.
AngleContrast an in-memory agent that loses its thread on redeploy with an externalized-state agent that binds on start; then argue that every framework adapter must satisfy a serialization contract, which is the real cost of the design and the part vendors omit.
Source video ↗Instrument Identity Before You Instrument Cost
Cost and effort attribution in multi-agent swarms is an infrastructure problem that must be solved at design time, because the primary key for attributing tokens is agent identity, and once agents are multiplexed across users and rooms that key cannot be reconstructed — making cost optimization and human-versus-AI contribution accounting permanently guesswork.
AngleShow why per-request telemetry is insufficient and per-agent identity-scoped telemetry is the minimum; demonstrate the retrofitting failure mode where a single shared credential destroys attribution for every past run.
Source video ↗Hydration Conformance Suite
beyond-evalsAt most two of three mainstream agent frameworks (Codex-, LangGraph-, Claude-Code-style) can restore full task state after a forced mid-conversation process kill when adapted to the same continuity interface, and the failures are attributable to framework-specific in-process state rather than to the persistence store.
Proof of conceptDefine a minimal continuity contract (persist conversation + in-flight task, hydrate on start, bind runtime). Wrap three frameworks behind it. Drive each through a multi-step task, kill the process at step 3, restart, and diff reconstructed task state against a ground-truth transcript.
MeasurementPass rate per framework over N restart trials; count of state fields lost per failure; time-to-resume after hydration; number of framework-specific shims required to satisfy the contract.
Source video ↗Broker Under Duress
newA central collaboration broker reduces lost messages versus point-to-point adapters under partial failure, but at the cost of a p95 hop latency increase large enough that synchronous agent-to-agent request/response patterns regress, while duplicate deliveries on retry remain nonzero unless adapters implement idempotency themselves.
Proof of conceptBuild an identical three-agent topology twice — direct adapters versus broker — and inject faults: dropped connections, restarts during in-flight tasks, downstream timeouts. Log every message with a monotonic id and compare delivery sets.
MeasurementOrdering violations per 10k messages, duplicate deliveries on retry, lost-message rate under injected faults, p95 and p99 hop latency with and without the broker.
Source video ↗Human Relay Accounting
movement-labHuman-mediated context transfer between parallel agent sessions accounts for a measurable double-digit fraction of wall-clock time and of total token spend in multi-session coding tasks, and eliminating the human relay reduces both more than improving the underlying agents does.
Proof of conceptInstrument a multi-session workflow: log every manual copy-paste/context re-injection event with its token size and elapsed time. Then replace the human relay with a scripted registration plus message-passing path and re-run the same task set.
MeasurementMinutes of human relay per task, tokens re-serialized per relay event, total tokens per completed task, error/retry count, before versus after the automated path.
Source video ↗Attribution Key Enforcement
gatehouseSystems that mint a per-agent identity principal (with delegating-human binding) before the first message retain full cost attribution across multiplexing, whereas systems starting from a shared credential cannot recover per-developer attribution for any historical run, only from the retrofit date forward.
Proof of conceptStand up two agent rooms on the same broker: one with per-agent principals plus delegation tokens, one with a shared credential. Replay an identical workload, then attempt to answer 'tokens spent per agent and per developer per task' for both.
MeasurementPercentage of historical runs attributable to an agent and a human; percentage attributable to a task; number of runs with ambiguous or colliding identity; audit-query latency.
Source video ↗