Open-source multi-agent workspace

Orchestrate the work.
Not just the model.

AgenZync turns AI agents into a coordinated, supervised team—so research, implementation, review, and specialist work can move in parallel without stepping on each other.

Discover the system See the orchestration loop ↓
MPL-2.0Local-firstDurable runsModel-flexibleHuman-supervised
mission / ship-auth-hardening live
ROOT ORCHESTRATORHarden authentication flow
3 agents active
AZ
R
Researchmapped attack surface
✓
I
Implementediting session.py
V
Verify12 checks queued
02
WRITE INTENTrevision 03
core/session.py · optimistic scope
implementation-agentoverlap tracked
TEAM-AWARE SKILLS✦SCOPED WRITE INTENTS✦DURABLE MISSIONS✦GUARDIAN APPROVALS✦REVISIONED JOBS✦MCP-READY✦EVIDENCE-BACKED AUDITS✦MODEL CHOICE
The shift

One agent can answer.
A system can deliver.

Today’s tools make individual agents faster. AgenZync is built for the harder problem: dividing consequential work, coordinating ownership, preserving context, and proving what the team actually did.

A single-agent loop
01Prompt
02One long context
03Hope it all lands cleanly
An AgenZync mission
Z
PlanBuildProve
✓ Scoped ownership✓ Conflict-safe writes✓ Reviewable evidence
The orchestration loop

One agent runs the team.
You stay in command.

The root agent does more than delegate. It follows every agent's status, messages, attempts, and outputs—then coordinates the next move, redirects work, or interrupts a run when the mission drifts.

ROOT / SHIP-AUTH-HARDENING supervising
AZ
MISSION OWNERCoordinate, inspect, and steer
ACTIVE
A1
Research agentattack-surface.md delivered
complete
A2
Implementation agentattempt 02 · editing session.py
attention
A3
Verification agentwaiting on changeset
queued
ROOT → A2

Redirect: preserve the public session contract, then send the changeset to verification.

MessageFollow upRedirectInterruptClose / reopen
01Observe

See durable threads, attempts, messages, and outputs across the team.

02Coordinate

Route context and follow-up work without losing the mission history.

03Steer

Redirect or interrupt an active agent when evidence changes the plan.

04Recover

Close, reopen, and continue work from inspectable state—not a black box.

COLLISION CONTROL

Same file.
Not the same conflict.

Agents execute in isolated Git worktrees. Scoped Write Intents make planned overlap visible, while submitted ChangeSets compare the actual line hunks. That keeps coordination precise without treating an entire file as one indivisible lock.

auth/session.py3 isolated worktrees · optimistic intents
A1
validate_token()L84—112
NO HUNK COLLISION
A2
rotate_session()L201—236
NO HUNK COLLISION
A3
validate_token()L101—124
RECONCILE

Same-path optimistic intents can coexist and stay visible. Real hunk overlap triggers re-read and explicit reconciliation; protected resources remain exclusive. Symbol evidence sharpens the report—it never weakens path and Git checks.

Modules with a point of view

Not every mission
is a coding mission.

Start from a focused Project, enable only the capabilities it needs, and give the team a skill package shaped around the domain—not a generic bag of prompts.

01

Code

Plan, implement, verify, review, and preserve the evidence behind every change.

implementationverificationpatch review
Explore Code ↗02

Audit

Split complex audits into source-backed claims, evidence work, and skeptical judgment.

orchestrationcitation proofevidence
Explore Audit ↗03

Creator

Move from voice and audience research to fact-checked, reusable content systems.

voicee-bookfact check
Explore Creator ↗04

Research

Run parallel investigations while keeping sources, artifacts, and conclusions traceable.

collectionsynthesisreview
Explore Research ↗05

Jobs

Turn recurring work into revisioned forms, skills, policies, and explicit output contracts.

typed inputsfrozen recipererun
Explore Jobs ↗06

Custom

Compose your own team, tools, policies, and deliverables around the work you actually do.

your rolesyour policyyour stack
Explore Custom ↗
Multi-agent skills + team-aware packages

Make the team part
of the skill.

An AgenZync skill can declare its multi-agent strategy and expected output in validated frontmatter. A required Multiagent Setup then describes when to delegate, how to divide the work, what workers return, and how the owner verifies and synthesizes the result. Packages can compose those skills into larger role and artifact recipes.

  • ✓ Validated strategy and output metadata
  • ✓ A required delegation and synthesis contract
  • ✓ Recommended roles, skills, and artifacts
  • ✓ Runtime-enforced depth and concurrency limits
citation-proof/SKILL.mdskill contract
---
name: citation-proof
multiagent: true
multiagent-strategy: supervisor-workers
multiagent-output: verified reference ledgers
  and a final citation-proof summary
---

## Multiagent Setup

- Coordinator inventories references and uses
- Workers verify independent source records
- Owner reads artifacts before synthesis
- Recover only the smallest failed unit
contract validruntime-guided delegation
Repeatable work, without prompt folklore

Define a Job once.
Run the workflow with fields.

Package a task template, validated inputs, skills, model and team policy, retry rules, and required artifacts into a revisioned Job Definition. Every launch pins the exact recipe it used.

Explore Jobs ↗
RUN / CITATION REVIEW
Durable by default

If it ran,
it left a trail.

AgenZync adds an orchestration-level record beyond the chat transcript. Observable execution state, events, agent activity, source material, artifacts, and reports persist on disk—ready to inspect during the run or revisit long after it ends.

runs / mission-042persisted workspace
⌄mission-042/live
↳events.jsonl248 events
↳runtime-context.jsonpolicy + model
↳state.sqlite3agents + leases
↳repo-manifest.jsonworkspace state
⌄worker-batches/3 agents
⌄fetched-sources/12 sources
↳final-report.mdready
↳quality-review.mdpassed
EVENT STREAM recording
intent

implementation-agent declared auth/session.py

agent

verification-agent started attempt 02

artifact

reports/auth-evidence.md persisted

review

Guardian requested human approval

Not private chain-of-thought.

A structured, inspectable record of the actions, state changes, tool results, handoffs, and evidence that make agent work accountable.

01Inspect live

See agents, events, leases, and progress while work is running.

02Resume safely

Recovery state and durable artifacts survive interrupted missions.

03Audit later

Trace claims and changes back to sources, actions, and review files.

Bring the agents you trust

Your models.
One operating layer.

Coordinate Codex CLI, Claude Code, OpenCode, OpenAI-compatible providers, or local Ollama models from one supervised workspace. Change the engine without rebuilding the way your team works.

AgenZZync
CodexClaudeOpenCodeOllamaOpenRouter

Product names and trademarks belong to their respective owners. Compatibility does not imply endorsement.

Supervised autonomy

Move faster.
Keep the evidence.

Every mission stays inspectable—from the team graph and event stream to source-backed reports, review artifacts, retries, and recovery. Humans remain part of the control loop.

01

Workspace boundaries

Agents operate inside the workspaces and run context you select.

02

Guarded changes

Patch-first edits and write-intent checks keep parallel work reviewable.

03

Policy-aware tools

Shell, network, and risky actions can be observed or held for approval.

04

Redacted events

Secret-looking values are removed from the execution stream.

05

Human control

Pause, steer, answer questions, approve, deny, and rescue stale work.

06

Durable recovery

Leases, heartbeats, retries, and retained artifacts support long-running missions.

ON THE ROADMAP

Code Guard

A built-in review layer for policy-aware code risk, dependency changes, and security-sensitive diffs—designed to meet every mission before it reaches “done.”

CG
01

Open source is not a feature.
It is the trust model.

AgenZync is licensed under MPL-2.0. Inspect the orchestration, understand the policies, extend the skills, and keep control of the workspace where your agents operate.

Inspect the runtimeExtend the packagesOwn the deployment
Built for teams of agents. Designed for the human in charge.

Bring ambitious work.
AgenZync the team.

The open-source coordination layer for AI work that needs more than one context window—and deserves more than blind trust.

Explore the architecture