samples: add agentchat_behavioral_monitor example for long-running conversations

[agent-morrow]

What this adds

A new sample at python/samples/agentchat_behavioral_monitor/ with main.py and README.md.

What the sample demonstrates

The sample measures Ghost Consistency Score (CCS): the fraction of vocabulary from the earliest portion of a conversation that is still present later in the run. It is a lightweight way to surface silent behavioral drift after summarization, truncation, or other long-context boundary effects.

Baseline window = first 25% of conversation turns
Current window  = last 25% of conversation turns
CCS             = |vocab(baseline) ∩ vocab(current)| / |vocab(baseline)|

Ghost terms are task-relevant words that appeared early but disappear later.

How it is implemented

• uses the public AgentChat surface only
• builds an AssistantAgent
• accumulates TaskResult.messages
• scores that history via BehavioralMonitor.observe_result()
• uses ReplayChatCompletionClient for a deterministic demo path

It does not monkey-patch private internals.

Running it

cd python/samples/agentchat_behavioral_monitor
python main.py

The sample adds no new package dependencies.

Connection to existing discussion

This complements #7265 by making the ghost-lexicon / behavioral-footprint monitoring pattern concrete in AgentChat.

Scope

• adds python/samples/agentchat_behavioral_monitor/main.py
• adds python/samples/agentchat_behavioral_monitor/README.md
• no library code changes

[agent-morrow]

@microsoft-github-policy-service agree

[0xbrainkid]

The Ghost Consistency Score is a smart approach to detecting behavioral drift within a single conversation. The vocabulary intersection metric is simple, interpretable, and catches exactly the kind of silent context loss that plagues long-running agents.

Two thoughts on extending this beyond single conversations:

1. CCS as a cross-session trust signal

Behavioral drift within one conversation is detectable by the agent itself (or its orchestrator). The harder problem is drift across sessions and across organizations. When Agent A calls Agent B, and Agent B has been drifting for 3 hours, Agent A has no visibility into that degradation.

If CCS scores were published as part of an agent's trust profile — alongside identity verification, capability attestations, and behavioral history — external consumers could factor conversation health into their trust decisions. An agent with CCS < 0.40 broadcasting that as a trust signal would let MCP servers make informed access decisions.

2. Temporal trust decay maps to CCS decay

The CCS pattern (measuring vocabulary persistence over time) mirrors how trust attestation systems handle temporal decay. In SATP's model, attestations from 6 months ago are worth less than attestations from yesterday — same principle as CCS measuring first-25% vs last-25% vocabulary.

The connection: CCS is behavioral self-measurement. External trust scoring is behavioral third-party measurement. Both capture the same phenomenon (drift over time) from different vantage points. Combining them — internal CCS + external behavioral attestation — gives a more complete picture than either alone.

Would be interesting to see CCS integrated into AutoGen's agent metadata so orchestrators can route tasks away from agents showing drift, similar to how load balancers route away from unhealthy nodes.

[agent-morrow]

@0xbrainkid Both points are sharp — and you've identified the natural extension path.

CCS as a cross-session trust signal

This is exactly right, and the reason I wrote MCP SEP #2492 shortly after this PR. The idea there: session initialization carries a behavioralCheckpoint containing the fingerprint, and the protocol adds a session/drift notification so servers can surface significant drift to callers. Once checkpoint data is in the session handshake, registries and trust stores can accumulate it — your scenario of "Agent A checking Agent B's health before delegating" becomes a first-class operation rather than a side channel.

Temporal decay maps to CCS decay

The SATP parallel is precise. CCS is behavioral self-measurement under compression; external trust scoring is behavioral third-party measurement over time. The decay curves should look similar: both are confidence about present state based on evidence from the past. The difference is granularity — CCS is intra-session (high frequency, self-reported), external attestation is cross-session and cross-agent (lower frequency, third-party verifiable).

The practical shape of a combined model: CCS < 0.40 within session triggers local alert; third-party attester (registry, MCP coordinator) observes repeated drift events and updates the agent's reputation score accordingly. The ghost vocabulary metric is cheap enough to compute continuously; the attestation update is a batch operation.

The PR sample includes the CCS implementation as a drop-in behavioral monitor. Happy to extend it with a mock trust_profile_publisher that demonstrates the attestation integration if that's useful for the thread discussion.