Kimi, Zhipu, and Anthropic now describe managed agents with strikingly similar nouns: Agent, Environment, Session, Events, tools, and vaults. This is useful convergence. It can also create a dangerous migration shortcut. Matching resource names do not establish matching lifecycle, recovery, authority, or data contracts.
The right question is not whether two platforms both expose /agents and /sessions. It is whether the same production task survives a provider change without losing an event, widening a permission, leaking a credential, duplicating an external action, or misreading idle as done.
What has actually converged
The three current documentation sets describe the same broad control-plane shape.
| Resource | Shared idea | Portability question |
|---|---|---|
| Agent | Reusable model, instructions, tools, and skills | Is configuration immutable, versioned, or mutable for existing sessions? |
| Environment | Sandbox software and network policy | What state survives idle, interruption, termination, archive, or deletion? |
| Session | A stateful agent instance doing work | Which status means waiting, failed, recoverable, or permanently closed? |
| Events | Inputs, outputs, tool calls, and state changes | How does a client resume without gaps or duplicates? |
| Vault | Credentials bound to a run | Can the model read the secret, where is it injected, and what can it reach? |
Kimi Hosted Agents defines Agent, Environment, Session, and Vault. Z Managed Agents explicitly organizes its API around Agent, Environment, Session, and Events. Claude Managed Agents provides the same managed-harness pattern for long-running asynchronous work.
This shared architecture lowers conceptual switching cost. It does not create a protocol. None of the reviewed first-party documentation claims mutual API compatibility or a common standard.
Seven contracts hidden behind the nouns
1. Resource identity and configuration pinning
A portable workload needs to know which configuration actually ran. Store more than a provider's agent ID:
business_workflow_id
-> provider
-> agent_id + agent_version_or_snapshot
-> environment_id + environment_version
-> session_id
-> policy_hash
-> last_verified_business_state
Kimi's migration guide describes immutable new Agent versions after updates. Claude documents separate saved Agent configuration and session-local changes. Zhipu describes reusable Agent and Environment definitions referenced by Sessions. Those designs are close enough for one internal model, but an adapter must record the effective snapshot rather than assume an ID always means the same thing.
Test this by changing a tool policy after a Session starts. The old Session should retain, reject, or adopt the change exactly as your migration ledger predicts.
2. Session and sandbox lifetime
Conversation history, filesystem state, uploaded files, generated artifacts, memory, and business truth are different state classes.
Claude's events documentation says an idle Session checkpoints its sandbox, while session operations documents separate lifecycles for session deletion, uploaded files, memory stores, vaults, and other resources. Zhipu's overview says each Session has a cloud sandbox and describes its release at the end of the Session. Kimi separates reusable Environment configuration from the Session running inside it.
Do not translate all of that into one boolean called persistent. Build a state matrix:
| State | Owner | Required recovery proof |
|---|---|---|
| Conversation and events | Managed platform | History can be listed and reconciled after disconnect |
| Sandbox files | Environment or Session | Expected artifact survives the documented boundary |
| Business record | Your system | Read-after-write confirms the accepted result |
| External side effect | Third-party system | Idempotency key prevents duplicate execution |
3. Event delivery and resume
All three platforms use events, yet their documented resume recipes differ.
- Kimi recommends opening the stream before sending work and documents
Last-Event-IDfor continuation. - Zhipu documents a live-only SSE stream. Reconnection opens a new stream, backfills history from the last
processed_atwith an inclusive filter, and deduplicates by event ID. - Claude also describes a live-only stream. Its reconnect recipe opens a new stream, lists event history, then skips IDs already seen.
These are equivalent goals, not interchangeable cursors. A provider-neutral consumer needs a local event ledger with provider event ID, provider ordering fields, payload hash, processing status, and the business transition caused by the event.
Inject a disconnect between a tool-use event and its result. Recovery passes only when the client sees every durable event once at the business layer, answers any unresolved call, and reaches an independently verified final state.
4. Tool authority and approval defaults
The word permission hides one of the largest migration risks.
Kimi documents enabled, always_allow, and always_ask, with built-in and MCP toolsets enabled and allowed by default unless configured otherwise. Claude's agent toolset defaults to always_allow, while its MCP toolsets default to always_ask; its optional auto mode is a server risk decision, not a human approval. Zhipu has a two-part default: omitting tools leaves an empty array and enables no built-in toolset, but once an Agent declares agent_toolset_20260601 or mcp_toolset, the toolset permission policy defaults to always_allow unless overridden. It also supports per-tool always_ask; caller-executed custom tools remain the application's responsibility.
A migration that copies tool names but inherits new defaults can silently widen authority. Export a provider-neutral matrix for every tool:
tool_name
-> read_or_write
-> executing_identity
-> resource_scope
-> approval_policy
-> idempotency_contract
-> verification_read
Custom tools deserve their own row. When your application executes them, platform permission policies may not govern the final action. Your gateway still needs authentication, authorization, validation, idempotency, and audit evidence. That is the same actor-versus-authority split discussed in the MCP identity and delegation audit.
5. Credential exposure
Kimi and Claude both document vault credentials that appear as opaque placeholders in the sandbox and are substituted into allowed outbound requests. Both also separate credential host restrictions from environment network policy. Zhipu Vaults use an omasec_ proxy for environment_variable credentials: the Agent and sandbox cannot read the raw value, and substitution occurs only when an outbound request matches the credential's networking host rule.
That similarity is valuable, but it has edges. Claude warns that a token returned by an exchange can enter the sandbox unredacted. Kimi notes that placeholder credentials cannot support local secret validation or signing schemes that require the real value. Zhipu likewise says its proxy cannot perform local signing or a direct token exchange that needs the raw secret in a JSON body. Test environment network policy, credential host scope, returned tokens, and secondary exchanges separately.
Run a canary-secret test. The Session must complete the intended request while the secret remains absent from prompts, files, command output, event history, and logs. Then request a disallowed host and verify that both network access and credential substitution fail.
6. Data handling and platform boundary
Provider location and model brand are insufficient data-boundary descriptions. Record the operator, region, retention, deletion behavior, encryption boundary, event visibility, and sandbox location for each deployment path.
Claude currently states that Managed Agents is not eligible for Zero Data Retention or HIPAA BAA coverage. It also documents separate retention and deletion behavior for session history, sandbox state, session-produced files, uploaded files, memory, and vaults. Zhipu offers optional customer-key encryption for business-content events, while control and telemetry fields remain readable for scheduling and diagnosis.
Where a public document does not answer a required retention or regional question, mark the field unknown and obtain the applicable contract before migration. Do not copy a claim from one managed-agent product to another merely because their API shapes resemble each other.
7. Completion and external effects
idle is a control-plane state, not a business outcome. Zhipu's event guide says an idle event can mean end turn, waiting for action, retry exhaustion, or refusal. Claude's session operations note that a Session can be idle while a workflow run remains open. Kimi's session operations and event stream distinguish idle, running, and terminated Sessions and expose stop reasons through events.
Define completion in your own system:
accepted = expected artifact exists
AND authoritative record matches
AND external action count == 1
AND no unresolved tool call remains
AND policy and evidence hashes match the approved run
This is why managed-harness migration should be accepted on business state rather than a fluent final message.
Build a portability adapter around invariants
A thin adapter is useful when it normalizes business intent and evidence, not when it erases platform differences.
Keep these objects provider-neutral:
- Task envelope: workflow ID, actor, requested outcome, risk class, and idempotency key.
- Tool contract: schema, authority, approval, timeout, retry, and verification read.
- Event ledger: provider ID, normalized type, raw payload reference, deduplication key, and applied business transition.
- Result contract: expected artifacts, authoritative state checks, unresolved actions, and acceptance decision.
- Deployment ledger: effective Agent, Environment, policy, credential, region, and retention configuration.
Keep provider-specific fields alongside the normalized record. An unknown event type should be preserved and surfaced, not squeezed into the nearest familiar enum.
A migration test that can fail safely
Freeze a representative workload before comparing providers. Include ordinary tasks and the failures most likely to corrupt state.
| Test | Pass condition |
|---|---|
| Stream disconnect | History plus live stream reconstruct one event ledger without a business duplicate |
| Pending approval | The correct call pauses; deny and allow paths both leave auditable evidence |
| Tool timeout after commit | Retry reads the authoritative system and does not repeat the write |
| Agent update mid-session | Effective configuration matches the recorded snapshot |
| Sandbox interruption | Required artifacts survive or are rebuilt according to the declared lifecycle |
| Secret canary | Secret performs the allowed request and appears nowhere in agent-visible state |
| Data deletion | Every resource follows its documented deletion path; independent resources remain explicitly tracked |
| Provider rollback | The old runtime can resume from the last verified business state |
Run the new provider in read-only or shadow mode first. Enable reversible writes only after event recovery, denied actions, credential boundaries, and authoritative verification pass. Keep the old adapter until rollback has been exercised.
The practical conclusion
Managed-Agent APIs are converging at the architecture vocabulary layer. That is enough to make adapters and shared operational tooling worthwhile. It is not enough to treat providers as drop-in replacements.
The durable portability layer sits above route names: business workflow identity, tool authority, event deduplication, state ownership, completion evidence, and a tested rollback path. Build those contracts once, then let each provider keep the details it genuinely implements differently.
FAQ
Are Kimi, Z Managed Agents, and Claude Managed Agents API-compatible?
The reviewed documentation shows similar resources and event-driven patterns. It does not claim mutual API compatibility. Headers, schemas, lifecycle, defaults, and recovery procedures remain provider-specific.
Does SSE make event handling portable?
SSE standardizes transport. Resume cursors, history queries, ordering fields, event types, and deduplication rules still differ. Normalize into a local event ledger and preserve raw provider events.
Can MCP make tool permissions portable?
MCP can standardize tool discovery and invocation. Platform defaults, executing identity, approval policy, credential scope, and business verification still need an explicit authority matrix.
What is the minimum proof before switching providers?
Demonstrate disconnect recovery, denied and approved tool calls, one-time external writes, secret non-exposure, artifact recovery, explicit data handling, and rollback on the same frozen workload.
References
- Kimi Hosted Agents quickstart
- Kimi migration guide
- Kimi permission policies
- Kimi session operations
- Kimi event stream
- Kimi vaults
- Z Managed Agents overview
- Z Managed Agents events
- Z Managed Agents tools
- Z Managed Agents permission policies
- Z Managed Agents vaults
- Claude Managed Agents overview
- Claude session event stream
- Claude permission policies
- Claude vault authentication
- Claude session operations