Skip to content

Orchestration Layer · Event bus

In development

Events, handoffs, webhooks and shared memory on one bus, with every run traced end to end. Agents that cooperate, fail loudly and never act on high-risk tools alone.

orchestration · run 7f3a · quote request
Event bus · live samplerequest.received
EVENT BUSIntake agentreads requestsTriage agentclassifiesQuote agentdrafts quotesCRM webhookinbound eventsShared memorycustomer contextHuman approvalvia Rendly

Trace · 8 spans

2,480 ms end to end

0 ms2,480 ms
  1. email.received

    40 ms
  2. intake.parse

    180 ms
  3. memory.lookup_customer

    50 ms
  4. triage.classify

    370 ms
  5. handoff → quote_agent

    28 ms
  6. quote.draft

    1242 ms
  7. risk.check send_quote

    36 ms
  8. approval.request

    534 ms

What it handles

Everything between one agent finishing and the next one starting.

Most agent projects fail in the gaps: a handoff nobody logged, a webhook that silently stopped, context that never arrived. The Orchestration Layer owns those gaps.

Events

Agents publish what happened. Others subscribe to what they need. No brittle point-to-point wiring.

quote.requested → triage_agent

Agent handoffs

One agent finishes its job and passes the work, with context, to the next. Every handoff is an event you can see.

triage_agent ⇢ quote_agent · ctx 3

Webhooks

Your CRM, ERP and inbox trigger agents, and agents call back out, through one governed entry point.

POST /hooks/crm · 200 · 84 ms

Shared memory

Customer context, prior decisions and state live in one place, so agents stop asking each other the same question.

customer:4471 → sla_tier, last_quote

End-to-end traces

Every run is traced from trigger to outcome, with each span timed. Know exactly where the seconds went.

run 7f3a · 8 spans · 2,480 ms

Error history

Every failure is logged with what happened next: retried, escalated or stopped. Nothing fails quietly.

quote.draft timeout → retry 1/3 ✓

End-to-end traces

One handoff, traced from trigger to human.

A quote request arrives by email. Three agents and a human touch it. The trace shows every event, every handoff and every millisecond, so when something is slow or wrong you know exactly where.

  • Agent spans
  • Tool and memory calls
  • Handoff on the bus
  • Waiting on a human

run 7f3a · quote request

8 spans · 2,480 ms · 1 handoff

Concept UI · sample data
0 ms2,480 ms
  1. email.received

    webhook

    40 ms
  2. intake.parse

    intake_agent

    180 ms
  3. memory.lookup_customer

    shared memory

    50 ms
  4. triage.classify

    triage_agent

    370 ms
  5. handoff → quote_agent

    bus

    28 ms · context: 3 items
  6. quote.draft

    quote_agent

    1242 ms
  7. risk.check send_quote

    policy

    36 ms · medium
  8. approval.request

    Rendly

    534 ms · waiting on a human
orchestration · error history
Today · 5 events0 silent failures
  • quote.draftrun 7f3a08:41:07

    Model timeout after 30 s

    → Retry 1/3 succeeded

  • crm.updaterun 81c009:12:55

    Webhook returned 401

    → Owner alerted in Rendly

  • intake.parserun 8a2210:03:19

    Attachment unreadable

    → Routed to a human

  • triage.classifyrun 9d0411:47:02

    Confidence 0.41 below 0.70 threshold

    → Held for human review

  • dispatch.planrun a1f913:30:44

    Auth token expired

    → Run stopped · logged

Logging and error history

Agents fail. Ours fail out loud.

The most expensive agent failure is the one nobody notices: a routine that stopped three weeks ago, a token that expired overnight. Every error is logged with what happened next, and the owner hears about it.

Retried

Transient errors retry with limits.

Escalated

Owners alerted in Rendly.

Stopped

Unsafe runs halt and are logged.

Principles

Built so agents can be trusted, a little at a time.

One agent, one job.

Small agents with one clear responsibility are easier to test, trust and replace. The bus lets them work as a team without becoming one tangled agent.

Staged autonomy.

Agents earn rights the way people learn to ride a bike: drafts first, one action next, full rights only after weeks of clean runs.

Humans on high-risk actions.

No agent moves money or signs a commitment without a named person approving it. That is enforced by the tool's risk rating, not by a prompt.

Tool risk table

Every tool an agent can call is risk-rated before it ships. Example below.

Concept UI · sample data
ToolRiskControl
search_knowledge_baseLowLog only
draft_replyLowLog only
update_ticket_statusMediumLogged · granted per stage
send_quoteMediumHuman review until trusted
issue_refundHighAlways human-approved
make_paymentHighAlways human-approved

Staged autonomy · the bike method

Nobody learns to ride without stabilisers.

Agents earn autonomy in five stages. Each one has a clear gate, and autonomy can be dialled back at any time. Step through it, or let it play.

AutonomyLowest

Stabilisers on: the agent drafts, a human sends.

The agent
Drafts replies, quotes and updates. Sends nothing.
The human
Reads every draft, edits, and sends it themselves.
Move on when
Drafts need few edits on real work.

Agents that work as a team. And answer to yours.

The Orchestration Layer is in development. If you are running more than one agent, or planning to, tell us where the handoffs hurt.