Work graph for coding agents
Continuity gives every run a living graph: the goal, the branch it opened, the evidence it produced, the debt it found, and the next task it can safely pick up.
Orientation beats drift
An unguided agent drifts across competing signals until a human redirects it.
The agent starts from a prompt, follows several fading traces toward side quests and stale context, then needs a human redirect before reaching partial output.
prompt agent side quest stale thread unfiled debt partial output human redirectThe agent follows whatever signal is loudest. It can self-steer into side quests until a human re-enters the loop.
A work graph connects the active goal to tasks, decisions, evidence, debt, and the next action.
The graph starts with the active goal, branches into a decision and implementation task, reconciles evidence and debt, then selects the next runnable task.
Continuity gives every run provenance and direction: where it came from, what changed, and the next safe move.
Compaction survival
A compressed chat thread loses nuance as it becomes a short summary.
Signals for the reason, edge case, blocker, and next task converge into a compact summary while several important details fade away.
full thread why edge case blocker? next? short summaryCompaction keeps the headline, but the small reasons vanish: why the agent moved, what edge case mattered, and where the next run should resume.
A durable graph route survives compaction.
A routed path connects goal, decision, evidence, current task, and next task through a compaction gate without losing provenance or direction.
The graph is not just prose in the chat window. A future agent can inherit provenance, blockers, evidence, and next direction after the thread is compacted.
Why Linear and Jira are not enough
Human-updated status fields drift behind the actual work.
The graph is reconciled from runs, deltas, evidence, blockers, and decisions.
Important reasoning sits in comments, Slack, or a forgotten agent transcript.
Context is typed: goals, tasks, questions, branches, debt, and proof stay connected.
A human still decides what an agent should read, trust, ignore, and do next.
The graph selects the next agent-safe slice and explains why other work is blocked.
Hard10 terminal-agent eval
Same task suite. Same terminal setting. The operating protocol changed the outcome.
Completed every run. Failed every hidden verifier.
The newer model ran faster and cheaper. It still solved zero.
Graph orientation, stop conditions, and reconciliation held across the full suite.
What changed
Model upgrades alone did not move the baseline: GPT-5.2 and GPT-5.5 both finished 10 clean trials and solved 0. With Continuity, GPT-5.2 solved all 10 by carrying graph state, blockers, and verification rules through the run.
Harbor provides the eval substrate. Terminal-Bench sets the standard for credible terminal-agent tasks. The full packet includes the Harbor runs, errors, costs, latency, and an interactive Codex TUI sidecar. This is a sharp signal, not the final product proof.
Open the full eval packet Read the trust protocolOutcome
The agent knows the outcome it is trying to create, not just the latest prompt.
Side quests become named future work instead of vanishing into chat history.
Each run records what changed, what was verified, and what still needs judgment.
The graph points the next agent at runnable work instead of another archaeology session.
trace exact work delta preserved
branch follow-up task created
exit clear next action
Pricing
Built for solo operators who want agents to keep shipping without rebuilding context every session.
Early access
Tell us where your agents lose the thread. The first cohort is for indie and solo developers who already feel the handoff problem.