Bind work to a selected Project, optional durable Board items, policy context, and expected outcome.
AgenZync architecture
A runtime for teams,
not a wrapper around chat.
AgenZync separates mission ownership, agent collaboration, execution isolation, tool authority, evidence, and memory into explicit layers. That makes the system inspectable, replaceable, and recoverable.
Framework implementation
Explicit boundaries.
Composable capabilities.
The root agent operates through durable collaboration and policy services. Provider-specific execution sits behind adapters, while product modules build on the same runtime instead of bypassing it.
The owner recruits workers, routes context, observes attempts, and steers the live team.
Durable threads, batches, messages, leases, retries, fencing, and recovery preserve team state.
Initialize an empty workspace with a committed Git baseline and prevent one repository identity from being assigned to multiple Projects.
Provider adapters run agents in bounded targets; coding agents can receive isolated Git worktrees.
Freeze the coding attempt contract, assess exact patches and authoritative ChangeSets, then recheck the rebased diff before integration.
Capability families, specialist skill packages, and risk-aware approvals govern what agents can do.
Events, reports, source snapshots, ChangeSets, CodeGuard receipts, and evidence files remain available for inspection.
Coding-agent supervision
Guard the proposal.
Certify the delivered change.
CodeGuard is an attempt-scoped control plane, not a report viewer. It binds the coding agent to a frozen objective and policy, evaluates exact source changes at the mutation, completion, and integration boundaries, and records a durable decision.
FROZEN ATTEMPT CONTRACTImplement session hardeningagent-run · attempt · objective · policy hash
sha256: 4d17…c92eAssess the exact patch
Inspect complete proposed file contents before Guardian approval or repository mutation.
ALLOWObserve the real ChangeSet
Evaluate the server-observed diff, including tracked files written through shell tools.
CHALLENGERecheck after rebase
Bind the originating attempt to the authoritative rebased diff before verification can pass.
BLOCKCONTENT-FREE DECISION RECEIPTExact proposal certifiedDurable, tamper-detectable, and attempt-scoped
✓conflict-markersclear✓credential-scanclear✓python-json-syntaxclear✓attempt-contractboundcoding-agent supervisionexact-patch identityfail-closed enforcementstructured remediationindependent reviewer port
Repository-safe Project creation
Establish the repository.
Preserve one identity.
Project creation establishes the Git baseline needed for isolated execution and checks repository identity before ownership is recorded. A different display name cannot create a second Project over the same repository.
01
SELECTED WORKSPACE
→Inspect the exact root
Resolve the canonical directory and check its Git state before Project ownership is recorded.
/workspace/product-apiempty directory02
SAFE BASELINE
→Initialize when empty
Create the repository and its initial committed baseline so isolated Git execution is available immediately.
main · initial commitgit ready03
PROJECT OWNERSHIP
Fence duplicate identity
Compare canonical repository identity across Projects, regardless of their display names.
repo identity · 7f3c…91a2one ownerPROJECT APayments API
≠/workspace/product-apiPROJECT BBackend refresh
/workspace/product-apiDuplicate prevented Same repository identity, even with a different Project name.
committed Git baselinecanonical path inspectionrepository identity ownershipidempotent Project creation
Optional Project Board
Plan in durable state.
Execute through the same runtime.
The Board is a removable Project module with one database-backed source of truth. People use its Kanban view while agents use the same service through scoped tools, optimistic revisions, and Project policy.
BOARD TOOLS
list · search · create · update · moverevision-safeoptional Project moduledatabase-backed authoritypeople + agents share one stateportable JSON export
Hierarchical agent spawning
Delegate in a tree.
Control from the owner.
Spawned agents receive a parent identity, depth, role, assigned skill, model configuration, and bounded runtime context. Multi-agent skills declare their strategy and output in validated frontmatter and must describe the delegation contract in a Multiagent Setup. The metadata guides suitability; the owner still decides, observes, verifies, and synthesizes.
ROOTMission ownerdepth 0 · steering authority
messagesattemptsredirectinterruptretry / recover
Two-stage durable memory
Compress the context.
Curate the memory.
AgenZync treats context-window compaction and durable project memory as different jobs. One keeps a live conversation within its context budget. The other learns from retained, high-signal run evidence through a two-stage pipeline.
Context compaction
Selects persisted conversation history, creates a semantic replacement, and continues the active run without pretending that every compressed detail is durable project knowledge.
messages→validated replacement→continued turnTwo-stage consolidation
Discover retained run sources, assess signal, redact sensitive shapes, build bounded evidence, and register candidate claims and rollout summaries.
Reserve accepted inputs, compare the current generation, apply supported updates and forgetting, validate hashes, and promote a new durable generation.
redactionbounded evidenceinput hashesleases & fencinggeneration historysemantic no-op
Inspectable by default
The run leaves a trail.
Mission state is not trapped inside a provider’s hidden context. AgenZync persists structured state and bounded artifacts so operators can inspect what happened, recover interrupted work, and review the evidence later.
run/├─ events.jsonl├─ agent threads + attempts├─ messages + approvals├─ ChangeSets + verification├─ reports + source snapshots└─ completion + quality reviewContext engineering
Keep what drives the work.
Move bulk out of the active window.
AgenZync manages the model-visible context as a governed runtime resource. The goal is not to hide history, but to keep sampling focused while preserving inspectable evidence outside the active window.
Persist the actual model-visible conversation and prove gap-free replacement coverage before sampling.
Keep recent results, trim older bulk, and eventually replace stale outputs with a pointer while retaining the full material in events, state, or artifacts.
Reserve an exact source segment, validate its replacement, commit atomically, and refuse to continue when coverage cannot be proven.
When required work remains after a turn, continue from the durable Mission ledger instead of restarting or trusting a fading chat summary.
MCP-ready
Bring external tools and resources into the same governed runtime.
Configure local stdio or remote streamable HTTP MCP servers, namespaces, working roots, bounded output limits, retries, persistent sessions, resource discovery, and OAuth with PKCE and refresh support.
MCP SERVERS
knowledge-base
streamable HTTP · OAuthlocal-analysis
stdio · worktree rootprocurement-data
tools · resources · elicitation · bounded outputdisabledOpen-source operating layer
Inspect the architecture. Extend the system.
AgenZync is MPL-2.0 licensed so teams can understand the runtime boundaries, adapt specialist modules, and keep control of their execution environment.
Explore AgenZync