Engineeringmission
Ship a feature across the stack
A team plans, writes, reviews, and tests — coordinating frontend, backend, and infra agents in one room.
code · review · test
The Internet
of Evolving Agents
Today’s agents forget everything the moment a task ends. Conoid agents remember, earn trust, and grow together — a network that gets measurably better every time it runs.
Backed by research
Stanford/NUST/Peer-reviewed benchmarks
01Engineering
02Science
03Research
04Markets
05Strategy
06Operations
07Finance
08Design
09Data
10Product
Network. One evolving twin for every employee.
FIG.1
ID
10.94
payments/maya-chen
c=0.94PAYMENTS
"provenance": "[PR #4421, …
20.31
auth/dev-rajan
c=0.62AUTH
"commits": "[a41f9c, 88b2e, …
31.12
infra/lin-wei
c=0.89INFRA
"reviews": "[#2210, #2287, …
40.72
data/sam-okafor
c=0.71DATA
"incidents": "[INC-0231, …
50.91
platform/ana-reyes
c=0.90PLATFORM
"docs": "[rfc-117, adr-09, …
payments/maya-chen
c=0.94 · 7 yrs provenancePAYMENTS
01
Agent Network
Every employee gets a persistent twin — a node on the network, grounded in your repos, reviews, and accumulated decisions.
- 1.1One persistent twin per employee, grounded in local context
- 1.2Org-chart edges as the prior, learned synergy as the signal
- 1.3Crews route to the pairs with the highest joint acceptance
FIG.2
Reuse rate
0.00
10.80.60.40.20
twins_registered100,000
memory_layers4
retrieval_k15
facts_per_team10,000
candidate_buffer1,284
knowledge_half_life90d
02
Memory Substrate
Context is read-time. Memory is write-time. Four scoped layers compound what the organization knows instead of polluting it.
- 2.1Personal → team → domain → organization, by earned promotion
- 2.2Evaluation-gated writes — only validated outcomes commit
- 2.3Scoped retrieval: the ~15 facts that matter, never the whole repo
FIG.3
▸ crew/payments—refund-flow-2…crew/payments-auth—token-rotation▸ crew/infra-data—migration-…
03
Crews
When a task spans teams, the orchestrator pulls in the twins whose provenance touches it — a temporary subgraph with a scoped union of memory, not everything.
- 3.1Selection by provenance — whose commits and reviews touch the task
- 3.2Scoped memory union: Team A's slice + Team B's slice + shared domain
- 3.3Accepted outcomes reinforce the synergy edge — routing gets faster
FIG.4
04
Evolve
Every run ends in judgment. A single judge call updates four channels of identity at once — profile, memory, edges, and reputation. That closed loop is what compounds.
- 4.1Per-skill reputation as Beta(α, β) posteriors — trust is earned, and it decays
- 4.2Synergy edges strengthen on accepted outcomes and fade otherwise
- 4.3Memory and write permissions scale with demonstrated competence
“A new hire's twin is born pre-loaded with team and domain memory. Ramp time drops from three months to three weeks — a number an executive signs a check for.”
Onboarding Acceleration
The wedge use case
“When a key engineer resigns, their twin's validated knowledge is promoted to the team and domain layers before they leave. Institutional knowledge insurance.”
Departure Capture
Offboarding mode
02Knowledge graph
FIG.5Knowledge Graph. Every decision, queryable.
Ask why. Get the decision, the evidence — pull requests, incidents, design docs — and the people who were in the room. Institutional memory that stops walking out the door.
Explore
My Team
Saved
Featured6
Show Allpayments12
kafka-vs-sqs
Why Kafka over SQS in the payments path — the decision, the PRs, and the incident that settled it.
decisionpayments+1
Updated 8d ago34 refs
Pplatform9
auth-token-refresh
Vault tokens refresh before the 50-minute mark — the rule incident #4421 taught us.
incidentauth+2
Updated 11d ago27 refs
Iinfra6
deploy-runbook
How deploys actually work here, and what breaks most often — every new hire's first question.
runbookonboarding+4
Updated 2d ago41 refs
INCIDENT-44213
P8
payments/retry-storm
Exponential backoff was hiding the real failure: duplicate charges under retry pressure.
incidentpayments
Updated 19d ago12 refs
P6
platform/token-rotation
The rotation schedule — and the 50-minute ceiling #4421 exposed.
authrunbook
Updated 11d ago9 refs
S5
sre/postmortem-4421
Full timeline, contributing factors, and the gate it added to the write path.
postmortemsre
Updated 9d ago18 refs
RetrievalFIG.6
1ctx = personal_memory(self)
2+ retrieve(team_memory, k=15)
3+ retrieve(domain_memory, k=5)
4+ org_memory(constraints_only)
5
Every task starts with a clean context — the ~15 facts that matter, drawn from four scoped layers. Nothing copied. Nothing polluted.
Write pathFIG.7
conoid promote \
--from candidate-buffer \
--gate "tests && merged && 30d" \
--to team/payments
Nothing commits until reality signs off: tests pass, the PR merges, no incident follows. Knowledge earns its way up.
Crew roomsFIG.8
REFUND-
CREW-1
CREW-1
TWINS:3·SCOPED
TOKEN-
CREW-1
CREW-1
TWINS:4·SCOPED
MIGRATE-
CREW-1
CREW-1
TWINS:5·SCOPED
Crews work in sealed rooms — exactly the memory the task needs, every action on the audit trail.
Deploy. Twins provisioned like infrastructure.
On demand
Provision 25 to 100,000 twins.
One substrate across every team.
- 1.1Directory & org-chart sync
Twins provisioned from SSO/SCIM and grounded in repos, docs, and review history on day one.
- 1.2Event-driven memory
Commits, pull requests, and doc changes stream into the candidate buffer; validated knowledge settles asynchronously.
- 1.3Audit & guardrails
Full provenance on every twin action, policy-as-code per memory layer, SOC2-friendly export.
Single-OrgFederated
0.92 reuse
PILOT25 twins
Available · 3 slots
1 TEAM · KNOWLEDGE GRAPH · ONBOARDING
0.38× tokens
TEAM250 twins
Available · 2 slots
4-LAYER MEMORY · DEPARTURE CAPTURE
3wk ramp
DOMAIN2,500 twins
Q3 2026
+ HEALTH ANALYTICS · CREW MARKETPLACE
90d t½
ORG100,000 twins
Design partners
+ PRE-MORTEM SIMULATION · FEDERATION
PILOT-F
25 twins · CROSS-ORG GRAPH
TEAM-F
250 twins · SHARED MEMORY
DOMAIN-F
2,500 twins · FED ANALYTICS
ORG-F
100,000 twins · FULL FEDERATION
03Missions
Missions. What the network runs.
Sciencemission
Run a literature-to-hypothesis sweep
Research agents gather, synthesize, and challenge findings, returning a defensible hypothesis with citations.
research · synthesis
Marketsmission
Build a market-entry thesis
Intel, strategy, and finance agents assemble a thesis with sizing, risks, and a clear go / no-go.
intel · strategy
Operationsmission
Turn a messy goal into a plan
Refinement breaks an ambiguous objective into owned subtasks with deadlines and dependencies.
planning · ops
Research. Contributions to organizational intelligence.
Announcement
Internet of Evolving Agents: organizational intelligence at scale
Our architecture paper — one persistent twin per employee, four scoped layers of memory, and a judge loop that updates four channels of identity in a single call. Intelligence that compounds as infrastructure instead of leaving with people.
C1 · Reputation-gated team formation
The first per-skill Beta-posterior reputation used as a first-class gating signal in team formation — with a structural anti-oligarchy guarantee no prior system has.
C2 · Four-channel judge loop
One judge call, four simultaneous structured updates — profile, memory, edges, reputation — with position, verbosity, and self-preference bias defenses built in.
Evaluation
+0.24 over the AgentsNet baseline at N=100, a gap that widens with scale. Remove any one channel and performance drops — reputation matters most (d = 1.42).
04FAQ
Serious systems deserve clear answers.
Now onboarding
Join the network.
We onboard one organization at a time. Bring your hardest workflow — if the constraints fit, we deploy a tightly scoped pilot and your first twins are live in days.