Concordex watches the synthetic behavior moving through your systems, learns the pattern of each subject, recognizes drift the moment it appears, and gates action when policy demands. The chain of what was seen and what followed is the codex your operators and auditors share.
# Trace the decision your agent made at 14:02 UTC $ concordex agents trace agt_7HxQ --at 14:02 → Agent support-triage-bot (v2.3.1) → Action refund.approve (amount=$840) → Evidence 3 retrieved docs · 2 tool calls · 1 policy check → Risk tags [financial-action, customer-pii] → Drift within band (p99 latency, refund-rate) # Export defensible audit packet for legal/compliance $ concordex audit export agt_7HxQ --window 24h --format soc2 ✓ packet written to audit-2026-05-07.zip (412 events, signed)
The refund agent was retrained on a Friday. By Monday, the breaker had stopped four attempts to write to a deprecated endpoint. The chain showed which fine-tune introduced the change. We rolled it back on Tuesday.
An auditor asked which manifest version produced a refund decision in March. The hash on the trace pulled the exact bytes from the canon library. The conversation took eight minutes.
A customer’s score crossed the policy threshold at 11:42. The breaker opened. Thirty further attempts on the account were declined before they reached the payment processor. Support had the context by lunch.
Anima maintains the live record of every subject in the workspace. Augur reasons over each new event. Concordia gates the actions that follow.
Capabilities, signature behaviors, modes that recur, risk tags, and the evidence behind each claim — updated as the subject moves.
Each forecast is calibrated against the subject’s historical record, scored, and reviewable. Augur reports when the distribution of decisions shifts — the shape of what the agent is doing, not the shape of the request volume.
Which agents may talk to which, what evidence flows between them, how disputes resolve. Policy diff history and rollback are first-class; your security team reads the same artifacts the runtime enforces.
One call per turn. The roster of participants travels with each event; the speaker is named explicitly.
from concordex import Concordex
cx = Concordex(api_key="ck_...")
with cx.conversation(participants=[
{"subject_id": "user:ws:bot", "role": "agent"},
{"subject_id": "user:ws:cust", "role": "customer"},
]) as conv:
conv.says("user:ws:cust", "I want a refund.")
conv.says("user:ws:bot", "I can help.")
One call, one cached read. The result names which policies fired and the ledger anchor that holds the transition.
with conv.guard("user:ws:bot", raise_on_open=True):
conv.tool_call("user:ws:bot",
"refund.issue",
{"amount": 9900})
Dashboard, Slack, CI gate, audit packet. The same chain answers every reader.
$ curl https://api.concordex.dev/v1/library/manifests/sha256:b1a5d... \
-H "X-Concordex-Key: ck_..."
{ "canon_id": "cn_core_concordex", "version": "1.0.0", ... }
The same record answers the engineer at the dashboard, the auditor at the export, the SRE in the incident channel, and the security officer at procurement. Each reads the chain in their own register.
LangChain, OpenAI Assistants, in-house, vendor-supplied — every agent appears in the same workspace, with a roster, a soul, and a chain. Standardize what production looks like while each team keeps building the way they prefer.
Each reasoning trace stamps the corpus snapshot it ran against; each policy decision anchors to the ledger. SOC 2, HIPAA, EU AI Act, and internal model-risk frameworks read the same artifacts the runtime enforces.
When the breaker fires, the incident channel receives the scope_ref, the policies that matched, the soul version at the moment, and the ledger anchor. Two clicks open the reproduction in the dashboard. PagerDuty, Opsgenie, and Slack receive the same payload.
Per-tenant data isolation, configurable retention, BYO-cloud deployment, and a policy layer expressed as code. The security review reads the same files the runtime enforces; self-hosted on your VPC remains available end-to-end.
Start on the free tier. Past it, each line on the bill maps to a meter you can read in the dashboard. Dashboard seats stay unlimited; the whole team reads from the same record.
Open a tenant, stream events, install the Concordex Core Canon, and read the chain. The free tier never expires; it becomes the bottom of every paid bill.
When usage crosses the free tier, you’re billed only on the meters you exceed. Volume commitments and reserved capacity reduce the unit price as the chain scales.
| Meter | Included on free tier | Rate past the floor | What it measures |
|---|---|---|---|
| Events ingressed storage-class meter | 100k / mo | $0.20 / 100k events | Every event the SDK streams or a connector fetches. SDK and connector paths meter identically. |
| Storage canonical bytes | 5 GB | $0.15 / GB-month | Canonical bytes of frames persisted in the chain. Retention on your plan governs expiration. |
| Ingress reasoning delegate runtime | 50k frames / mo | $0.40 / 1k frames | The shaping pass that turns RawInput into a BehaviorFrame — runs the delegate, normalizes fields, attaches the subject id. |
| Corpus reasoning deep-model evaluation | 10k passes / mo | From $0.60 / 1k passes | Deep-model evaluation of a frame against your corpus. Tokenization is the cost driver — a larger corpus reads more vocabulary per pass. |
| Canon royalties pass-through | Core Canon | Set by author | Authors price each canon they publish — some per install, some per reasoning pass against their bytes. Visible on the canon’s Library page before you commit. |
Larger corpora yield sharper reasoning and cost more per pass. The Corpus page in the dashboard previews the per-pass token impact of every install before you commit.
Annual commitments on any meter reduce the unit price. Reserved-capacity tiers prepay a corridor at a discount. BYO-cloud and self-hosted deployments share the same control plane — your KMS keys, your VPC, the same ledger format. DPA, BAA, and a solutions architect on the account.
A tenant provisions in under a minute. Most teams send
production events the same afternoon — via the SDK
(pip install concordex /
npm i @concordex/sdk) or a sidecar proxy when the
runtime is opaque.
SaaS tenants run in us-east-1 or
eu-west-1 with configurable data residency.
Self-hosted deployment on your VPC is available on Enterprise
— same control plane, your KMS keys, traffic stays in
your VPC.
SOC 2 Type II is in flight (target Q3 2026); evidence available under NDA today. HIPAA-compliant deployment with BAA on Enterprise. Pre-mapped controls for EU AI Act Article 12 (record-keeping) and Article 14 (human oversight) ship in the export packet; the crosswalk is on the security page.
Concordex emits OpenTelemetry, ships first-party integrations with Datadog and Splunk, and webhooks for the rest. APM handles latency, errors, throughput; Concordex handles agent decisions, evidence chains, and risk tags. The two pair cleanly.
Every event is exportable to JSONL or Parquet at any time on demand. Self-hosted customers keep running on the last released image; the registry schema is documented and open.
The platform indexes how trust between many minds — human and synthetic — is built, sustained, and verified. The full thesis is on the thesis page.
A tenant provisions in a minute. First event inside five. First audit packet by the end of the week.