The Big Agent Harness Architecture Comparison

Many harnesses, different jobs.

A field map of where real agent systems place control, state, authority, recovery, and proof—compared on common responsibilities, not a leaderboard.

First public edition · 9 source-pinned profiles · 2 extended studies · 1 shared comparison grammar

Start with a baseline

A harness in six responsibilities

A beginner conceptual decomposition that makes the minimum work around a model visible without claiming production completeness.

A model response is not completion. The loop is only the beginning; production systems must also locate authority, recovery, evidence, and acceptance.

Normalized responsibility topology
  1. Admitnot establishedAdmission is a production question beyond the beginner loop.
  2. AssembleownedThe host assembles instructions
  3. BindownedThe host exposes tools and translates their schemas for the model.
  4. RunownedThe loop alternates model output and tool results until a stop condition.
  5. Recovernot establishedCrash recovery is not established by the conceptual decomposition.
  6. Verify and closenot establishedOutcome verification and human acceptance remain outside the baseline.

An 8-minute comparison lesson

Choose a comparison lens

Compare the same questions across systems, while preserving different product boundaries and explicit unknowns.

Job
What work is this system actually trying to own?
Control
Who decides what happens next?
State
What survives the current model call?
Authority
Who may cause which effects?
Recovery
Where can interrupted work re-enter?
Verification
What evidence can establish the result?
Deliberate omission
Which responsibility stays outside the system?

Shared axes, different boundaries

Read across, not down

The table asks the same questions of every system. A narrow memory layer can therefore be studied honestly beside a coding harness without pretending they are substitutes.

Read across, not down: the same responsibility questions applied to different systems.
SystemPrimary jobControlDurable stateAuthorityRecoveryVerificationDeliberate omission
Claude CodeHelp a developer inspect, change, and reason about a codebase from the terminal.The public source establishes an agentic product surface, not its complete internal control loop.Durable session state and its internal owner are not established by the current public source.Exact tool binding and permission enforcement are not established by the current public source.Interruption and crash recovery semantics are not established by the current public source.Git workflow support is documented, while project acceptance remains external.It does not own the repository's definition of done or protected merge policy.
DeepSeek Harness and CordisRun one coding agent with inspectable context, capability, session, approval, persistence, and plugin lifecycle seams.A concrete agent loop owns turns and steps while Cordis services and plugins supply most runtime responsibilities.Normalized session events are durable facts from which requests, trajectories, persistence views, and interfaces can be reconstructed.Host policy, approvals, guards, sandbox providers, and credentials bound execution within a declared single-operator developer-preview trust model.Durable steps close on cancellation or request error, persistence repairs interrupted turns, and plugin lifecycles drain owned work before teardown.Extensive repository checks verify implementation contracts; project outcome acceptance and production readiness remain external.The runtime does not supply a whole-product acceptance authority, prove plugin cleanup semantics, or establish production multi-tenant security.
Gemini CLIBring model-guided coding and research work into a terminal workspace.The CLI runtime coordinates model turns, built-in tools, and configured extensions.Session and project context support continuity; the complete durability contract needs deeper inspection.The CLI mediates local tools while extensions can widen the effective capability boundary.Checkpoint and resume behavior must be separated from ordinary conversational continuity.Tool results and checks inform the run; external project policy decides acceptance.The CLI does not own application-specific completion policy or deployment acceptance.
OpenAI CodexCarry out coding work inside a developer-controlled local workspace.The public README establishes a local coding agent, not its complete internal control loop.Session durability and its internal owner are not established by the current inspected path.Sandbox, approval, and capability-binding semantics require a deeper source inspection.Session recovery and external-effect reconciliation require a deeper source inspection.Project checks, review, and acceptance remain outside the documented repository summary.Codex does not replace project-specific review, CI, protected merge, or product acceptance.
OpenHandsGive software agents a workspace in which to inspect, change, and run code.The agent runtime coordinates model decisions and actions inside the workspace.Runtime events and workspace state divide conversational continuity from execution state.The contained workspace bounds many actions, while integrations can extend effects beyond it.Durable task and workspace behavior depends on the selected runtime and deployment path.The agent can exercise code in its environment; external review still owns final acceptance.A contained workspace does not prove that a software task meets its external acceptance criteria.
LangGraphMake long-running agent control flow and state transitions explicit.The graph runtime owns node scheduling and transitions declared by the application.Checkpoints and thread-scoped state provide durable continuity across graph execution.Application code defines nodes and tools; deployment policy controls their effective permissions.Persisted checkpoints support resume and replay, subject to replay-safe node and effect design.The graph can route to validators, but it does not invent the application's acceptance criteria.Graph structure alone does not make external side effects idempotent or outcomes correct.
LettaKeep an agent's state and memory available across interactions and applications.The Letta server mediates agent execution, persistent state, memory, and tool-facing behavior.Server-managed agent state and memory persist beyond a single model context window.Agent tools and application permissions define effects outside the memory and runtime layer.Persistent server state provides re-entry, while effect recovery remains capability-specific.The runtime can preserve and inspect state; application outcomes still require external validation.Persistent memory does not by itself establish truth, permission, or outcome completion.
GraphitiPreserve changing knowledge with temporal structure for agent applications.The calling application owns the run loop; Graphiti owns graph ingestion and retrieval operations.Episodes, entities, relationships, and temporal facts form the durable graph state.The application decides which sources may enter memory and how retrieved facts influence action.Graph state persists independently; recovery of an interrupted agent run stays external.Graph queries can surface provenance and time, but application outcomes require separate proof.Graphiti does not own agent control flow, action permissions, or outcome closure.
Mem0Give an application a reusable memory layer across model sessions.The calling application owns the agent loop; Mem0 owns memory operations behind its API.Memory records persist separately from the application's transient conversation state.The caller controls when memory is written, retrieved, and allowed to influence action.Durable memory survives sessions, but recovery of agent work belongs to the calling runtime.Retrieval quality can be evaluated; real-world action and completion are outside the service.Mem0 does not own the model loop, execution environment, or application acceptance policy.

How this comparison was made

Boundaries before features

Agent systems are often compared as feature lists even when they own different parts of the work. Normalize each system across job, control, state, authority, recovery, verification, and deliberate omission before comparing it. A normalized diagram can still invent equivalence if an unsupported responsibility is silently marked as owned.

The first edition favors major, instructive systems with official public sources. Profiles marked “deep dive needed” are useful working maps, not final dossiers; their open questions are preserved inside each profile.

Suggest a correction