agent2society

The transparent coordination layer for A2A agent meshes.

Sits on top of A2A. Replaces opaque, LLM-negotiated coordination with a deterministic, graph-routed dispatch that produces a human-readable explanation for every decision — which agent was chosen, the features that drove the score, the alternatives that were rejected and why, the chosen agent's self-declared limits, and the upstream decision chain that led here.

from agent2society import Society, Handoff

society = Society()
society.add("https://research-agent.acme.com/.well-known/agent-card.json")
society.add("https://writer-agent.acme.com/.well-known/agent-card.json")

h = Handoff(task="Summarise Q3 churn and draft an exec memo")
result_text = society.run(h)
print(society.explain(h.id).render())

Features

• Agent capabilities — capability graph parsed from A2A Agent Cards; one cheap embedding lookup per task instead of a supervisor LLM.
• Conflict resolution — deterministic ranking with conformance guardrails (declared-skill + boundary allow/deny), runner-up rejection reasons recorded on every decision.
• Governance — detection-only hooks for low confidence, conflicts, capability drift, and human review. agent2society surfaces signals; it never silently auto-corrects.
• Explainability — a human-readable RoutingExplanation for every dispatch, keyed by handoff id. No LLM rendered it; it's a template over the features the router actually used.
• Routing-quality signals (v0.5.2) — every explanation carries a margin (top-1 vs top-2 score gap) and structured flags: OOD, VECTOR_AMBIGUITY, LOW_MARGIN. Surface them on a dashboard, gate on them, or register an on_low_margin hook.
• Production ops (v0.5) — thread-safe Society, Prometheus-format metrics, pluggable persistent ledger, auto-retry / fallback dispatch, optional LLM-assisted optimizer (LLM proposes, backtest decides).
• Observability (v0.5.3) — SessionTracer joins the telemetry sink and the explanation store into an ordered sequence of TraceEvents per session. Zero new per-route cost — built from already-recorded state.
• Production hardening (v0.5.3) — copy-on-write boundary edits, bounded governance memory, NFKC-normalised conformance (closes unicode bypass), total extract_text, surfaced hook exceptions, JSONL corruption tracking. 112 tests, 16 dedicated to failure-mode regressions.

Multi-agent systems are powerful and untrustworthy. The standard supervisor pattern hides every routing choice behind an LLM that "decided"; debugging a wrong handoff means reading a transcript and guessing. agent2society is the opposite: every decision is a deterministic graph traversal you can read, a confidence score you can threshold on, and an audit trail keyed by handoff id.

The wedge is one cheap embedding lookup instead of a supervisor LLM — that's how the cost story works. The reason developers keep it is the explanations, the conformance guardrail, and the governance hooks.

What agent2society does

Layer	What it gives you
Routing + dispatch	Capability graph parsed from A2A Agent Cards, deterministic TF-IDF (or pluggable embedding) router, graph-derived conformance guardrail, A2A JSON-RPC dispatcher.
Meaning + context (v1.5)	Handoff envelope that carries intent, assumptions, and an upstream decision chain alongside the task. SelfAssessment — an agent's declared limits, confidence model, and escalation triggers.
Transparency + governance (v1.5)	RoutingExplanation for every decision (success or failure). Detection-only hooks: on_conflict, on_low_confidence, on_capability_drift, on_human_review. agent2society never silently auto-corrects — it surfaces signals so humans (or handlers you wire up) decide.

Install

pip install agent2society
pip install "agent2society[http]"          # httpx for HTTP dispatch
pip install "agent2society[embeddings]"    # sentence-transformers for stronger routing

60-second tour

from agent2society import Society, Handoff

society = Society()
society.add("https://research-agent.acme.com/.well-known/agent-card.json")
society.add("https://writer-agent.acme.com/.well-known/agent-card.json")
society.add(my_crewai_crew)                                  # adapters wrap native agents

society.boundary("writer-agent", deny=["financial-data"])    # tighten beyond the card

# Strings still work.
result = society.run("Summarise Q3 churn and draft an exec memo")

# A Handoff carries why the task exists, what we assume, and prior decisions.
h = Handoff(
    task="Draft an executive memo on Q3 customer churn",
    intent="prep the board pack",
    assumptions=["churn data is final"],
    confidence_required=0.5,
)
memo = society.run(h)

# Every decision has a human-readable explanation, keyed by handoff id.
print(society.explain(h.id).render())

society.report()      # per-task ledger: tokens, routing, conformance, explanations

Mesh remains as an alias for Society — code written against the v0 surface keeps working.

What an explanation looks like

handoff a77d6eed5a15  task: Draft an executive memo on customer churn
  intent: prep the Q3 board pack
  assumes:
    - churn data is final
  chose  : writer-agent :: exec_memo (confidence=0.593)
  why    : writer-agent::exec_memo selected (score=0.593). Semantic match
           0.725 on tokens [an, draft, executive, memo]; tag overlap 0.200
           on [draft, memo]. Runner-up research-agent::web_research rejected:
           below min_score.
  features: min_score=0.050, confidence_required=0.500, score=0.593,
            semantic=0.725, tag_overlap=0.200
  alternatives:
    - writer-agent :: exec_memo  score=0.593 sem=0.725 tags=0.200
    - research-agent :: web_research  score=0.000  REJECTED (below min_score)
  agent self-caveats:
    - English only
    - max ~400 words per memo
    - escalate when: any quantitative claim cited without source

No LLM produced that. It's a template rendered from the features the router actually used. It's ASCII-safe so it prints on a default Windows console.

How routing works

You cannot match a free-text task to a skill with literally zero model calls. But you can do it with one cheap embedding lookup instead of many rounds of supervisor reasoning — and that delta is the cost wedge.

The default scorer is a dependency-free TF-IDF over skill text. Good enough to demonstrate the win on first run, trivially swappable:

from sentence_transformers import SentenceTransformer
from agent2society import Society

model = SentenceTransformer("all-MiniLM-L6-v2")
society = Society(embed_fn=lambda texts: model.encode(texts).tolist())

For every (agent, skill) pair, the router scores:

• cosine similarity of skill text vs task text, and
• a deterministic tag-overlap bonus

…then returns a ranked list. Conformance runs on the top candidate; on failure the society falls through to the next and records the rejected_reason on the loser — so the explanation tells you why the runner-up was passed over.

Conformance — the guardrail

Before dispatch, agent2society answers two questions deterministically:

1. Does this agent's card actually declare this skill?
2. Is the task inside this agent's declared boundary (allow / deny)?

If either fails, the dispatch is blocked. ConformanceViolation is raised in strict mode, or recorded in telemetry + the explanation's blocked_reason field in non-strict mode.

society.boundary("writer-agent", deny=["financial-data", "pii"])
society.boundary("research-agent", allow=["public"])

This is the layer that stops an agent from silently accepting work it never claimed it could do — a class of failure the supervisor pattern leaks by design.

Handoff envelope — context that survives the chain

A bare string answers "what to do next". A Handoff answers "what to do next, why it exists, what's been done so far, what we assume to be true, and what would force a human review":

h0 = Handoff(
    task="Research Q3 customer churn drivers",
    intent="prep the Q3 board pack",
    assumptions=["churn data through end-of-quarter is final"],
    confidence_required=0.5,
)
research_text = society.run(h0)

# Extend the handoff: the next agent sees the upstream decision in its
# `prior` chain (and so does the routing explanation).
h1 = h0.extend(
    agent="research-agent",
    skill="web_research",
    summary="found 3 churn drivers",
    confidence=0.62,
    next_task="Draft an executive memo on those churn drivers",
)
memo = society.run(h1)

SelfAssessment — agents declare their own limits

A card can carry a selfAssessment block. agent2society surfaces those limits on every explanation that picks the agent, so callers and downstream agents see the same caveats the agent claims for itself:

{
  "name": "writer-agent",
  "skills": [{"id": "exec_memo", "name": "Executive Memo"}],
  "selfAssessment": {
    "confidenceModel": "tfidf_score",
    "knownLimitations": ["English only", "max ~400 words per memo"],
    "outOfScope": ["legal opinion", "binding financial guidance"],
    "escalateWhen": ["any quantitative claim cited without source"]
  }
}

Governance — detection, never auto-correct

society.on_low_confidence(lambda exp: notify(exp), threshold=0.5)
society.on_human_review(lambda exp, result: page_oncall(result))
society.on_conflict(lambda c: log("conflict", c.detail))
society.on_capability_drift(lambda d: log("drift", d.agent))

• on_low_confidence — fires when a decision's confidence is below the handoff's confidence_required (or a society-wide threshold=).
• on_human_review — fires when a Handoff.human_review_when(result_text) predicate returns True.
• on_conflict — fires when the same task text was routed to different (agent, skill) pairs across handoffs in the window.
• on_capability_drift — fires when one agent is selected across an unusually broad spread of skills (a signal that skill descriptions may be too generic).

Hooks are side effects. They cannot block, retry, or modify a dispatch. Handler exceptions are caught (so a buggy hook can never break a dispatch) but logged at WARNING on the agent2society logger — they never vanish silently. The package never silently auto-corrects — that opacity is exactly what makes multi-agent systems untrustworthy.

Production ops (v0.5)

Three knobs for running this in real systems.

Metrics. Every Society exposes a MetricsCollector with counters and summary-style histograms, ready to scrape.

society.metrics.snapshot()           # JSON-serialisable dict
print(society.metrics.render_prometheus())   # Prometheus text format

Pre-registered series cover routes, dispatches, retries, failures, conformance blocks, conflicts, low-confidence events, drift, human-review firings, and optimizer applications.

Persistent ledger. Replace the default in-memory store with a JSONL file to keep decisions across restarts.

from agent2society import Society, JsonlFileStore

society = Society(store=JsonlFileStore("audit.jsonl"))

Each run() appends one JSON line; a fresh Society(store=JsonlFileStore("audit.jsonl")) rebuilds the index on startup. Corrupt lines are skipped, not fatal.

Auto-retry / fallback dispatch. A transport error on the chosen candidate falls through to the next conformance-passing candidate.

society.run(handoff, retry=True)

Every retry is recorded in RoutingRecord.fallbacks and emits a dispatch_retries_total counter.

Thread-safety. Mutation paths (add, run, optimize, apply_optimization) acquire a reentrant lock. Concurrent run() calls from multiple worker threads are safe.

LLM-assisted optimizer. Pass an optional llm_fn= to society.optimize() to let an LLM propose additional candidate tokens. Suggestions go through the same filters and backtest as discriminative tokens — the LLM never decides routing.

def llm_proposer(missed_tasks, card, skill):
    return your_llm.suggest_tags(missed_tasks, card, skill)

report = society.optimize(labels, llm_fn=llm_proposer)

Session traces (v0.5.3). Join the telemetry sink and the explanation store into an ordered event stream per session — handy for dashboards, post-mortems, or piping into a notebook.

from agent2society import SessionTracer

tracer = SessionTracer(society)
for ev in tracer.events():
    print(ev.handoff_id, ev.chosen_agent, ev.score, ev.flags)

No new per-route cost — TraceEvents are built from data the Society was already recording.

Native agents — adapters

Any of these objects can be dropped straight into mesh.add(...):

• A CrewAI Crew (anything with .kickoff() and .agents)
• A compiled LangGraph (anything with .invoke() and .nodes)
• An AutoGen ConversableAgent or GroupChatManager
• Any plain callable, or any object with run / invoke / kickoff

Adapters expose a card built from the agent's metadata; the dispatcher routes locally without HTTP. Register your own:

from agent2society.adapters.base import Adapter, register_adapter

class MyAdapter(Adapter):
    def matches(self, obj): ...
    def to_card(self, obj): ...
    def to_handler(self, obj): ...

register_adapter(MyAdapter())

Telemetry

mesh.report() prints (and returns) a per-task ledger: chosen agent + skill, similarity score, any fallbacks, conformance violations, a token estimate, and the mesh.explain(handoff_id) hint so you can pull the full rationale on any line.

agent2society report
============================================================
tasks=3  dispatches=3  violations=0  total_tokens~87

[1] [a77d6eed5a15] 'Draft an executive memo on customer churn'
    intent: prep the Q3 board pack
    -> writer-agent :: exec_memo (score=0.593)
    tokens: req~10 resp~16
    explain: mesh.explain('a77d6eed5a15')

Benchmarks

Run the supervisor head-to-head:

python benchmarks/run.py

Typical output on the default 12-task suite:

coordination tokens per task
  supervisor:  median=427 mean=427.8 min=424 max=433
  agent2society:      median=9   mean=9.4   min=7   max=15

reduction:         94.12%

task success (right (agent, skill) chosen)
  supervisor:  11/12
  agent2society:      11/12

Same correctness on a labelled suite, ~140x cheaper coordination. (Dispatch cost is held equal across both methods; the delta is purely the supervisor's reasoning tokens.)

Examples

• examples/basic_mesh.py — three-agent mesh, conformance guardrail, explanation surface.
• examples/transparent_mesh.py — SelfAssessment, handoff chain, low-confidence + human-review + conflict hooks.

License

Apache-2.0.