Seamless Background-to-Foreground Handoff¶

Enable humans to take over from background agents at the ~90% completion mark — using distilled context summaries and durable artifacts — rather than cold-starting from raw output or waiting for full automation to reach 100%.

The Problem¶

Background agents running autonomously handle well-defined work efficiently, but most complex tasks have a nuanced tail: the final 10% requires judgment, aesthetic evaluation, or domain knowledge the agent cannot reliably supply. Without a structured handoff, humans face two bad options:

Cold start: pick up from artifacts (draft PR, file diffs) with no context on what the agent tried, what failed, or why decisions were made
Over-automation: let the agent push to 100%, accepting errors on the nuanced tail, then fix them in review

The nibzard/awesome-agentic-patterns catalog names this the seamless background-to-foreground handoff, attributed to Aman Sanger (Cursor): "quickly move between the background and the foreground is really important."

The Handoff Design¶

Four components make this work:

1. Completion Threshold¶

The agent stops before nuanced judgment calls — not at failure and not at 100%. The ~90% figure is illustrative; the real criterion is "the remaining work requires human judgment." Triggering points include:

Confidence falls below a threshold on a decision with external impact
Multiple valid options exist and the choice reflects preference, not correctness
The task requires testing against infrastructure or environments the agent cannot reach

2. Context Preservation via Distilled Summaries¶

The handoff passes a compact summary, not the full conversation history. The nibzard/awesome-agentic-patterns catalog describes significant compression between the agent's full execution trace and the handoff summary. The summary captures:

What was completed and what remains
Decisions made and alternatives rejected
Blockers and open questions for the human

Anthropic's Effective Context Engineering for AI Agents describes the same mechanism for agent context resets — compaction and structured note-taking (e.g., NOTES.md) — directly applicable to human handoffs. See Context Engineering for the broader discipline.

3. Artifact-Based Handoff Points¶

The handoff medium is a durable artifact the human can pick up in their native tooling:

Draft PR on a named branch — cloneable, openable in any IDE, independently continuable
Progress file (e.g., claude-progress.txt) — records what the agent completed and what it was about to do next
Git history — every commit the agent made is part of the handoff; the human reads diffs, not agent logs

Anthropic's Effective Harnesses for Long-Running Agents describes claude-progress.txt plus git history as the artifact pair for session resumption — the same mechanism works for human takeover.

4. Tool Parity¶

The human picks up using the same tools the agent used: same IDE, same terminal commands, same MCP servers. When tool interfaces match, the human can apply the agent's context summary directly — no translation of "what the agent did" into "what I can do." Tool parity reduces the impedance mismatch between agent execution and human continuation — the human operates in the same environment the agent described in its summary.

Progress Visibility¶

Real-time progress visibility changes when and how the human engages. Without it, the human either polls for completion or misses the handoff window. The relevant property: the human sees incremental agent output and can intervene before the agent reaches its stopping point, not just after.

Handoff Flow¶

graph TD
    A[Task assigned to agent] --> B[Agent executes autonomously]
    B --> C{~90% threshold reached?}
    C -- No --> B
    C -- Yes --> D[Agent compresses context to summary]
    D --> E[Agent commits progress artifact]
    E --> F[Agent opens draft PR / stops]
    F --> G[Human notified via progress stream]
    G --> H[Human reads summary + diffs]
    H --> I[Human continues in same tools]
    I --> J[Task complete]

This pattern is often conflated with two others:

Pattern	Trigger	Direction	Purpose
Human-in-the-Loop	Before irreversible action	Agent pauses mid-task	Gate to prevent errors
Cloud-Local Agent Handoff	Surface transition	Environment switch	Continue agent work on different infra
Background-to-Foreground Handoff	Near-complete task	Agent to human	Human completes nuanced tail

HITL gates interrupt the agent pipeline at risk points. Cloud-local handoff moves work across execution surfaces. Background-to-foreground handoff transfers ownership from agent to human at a planned completion threshold.

Example¶

A background agent is assigned: "Implement the pagination component from the design spec." After 45 minutes, it has:

Implemented all page navigation logic and keyboard shortcuts
Written unit tests (all passing)
Opened draft PR feat/pagination-component with 12 commits

The remaining work: visual polish that requires subjective judgment on spacing and animation timing that the spec leaves ambiguous.

The agent stops, writes a summary to claude-progress.txt:

Completed: Core pagination logic, keyboard nav, unit tests (23 passing).
Remaining: Animation timing on page transitions (spec says "smooth" — no ms value).
           Hover state color — design system has two candidates, neither specified.
Open question: Should pagination reset on filter change? Product decision needed.
Draft PR: feat/pagination-component (branch: feat/pagination-component)

The developer opens the branch in their IDE, reads the summary, makes three targeted edits, and merges.

When This Backfires¶

The pattern assumes the agent can reliably identify the 90% threshold and produce an accurate handoff summary. Both assumptions can fail:

No reliable stopping signal: agents that lack explicit completion criteria or confidence thresholds stop at arbitrary points — sometimes too early (wasting the handoff), sometimes too late (after the agent has already made the irreversible judgment calls the pattern was meant to preserve for the human).
Summary drift: if the agent's summary omits or misrepresents decisions made during execution, the human picks up from a false starting point. A draft PR with misleading context can be harder to untangle than starting fresh from requirements.
Tool parity absent: when the human's environment differs from the agent's — different branch state, missing MCP servers, inaccessible infrastructure — the handoff degrades to a cold start regardless of summary quality.

In these conditions, a hard 100% automation loop with post-hoc review is often more reliable than a mid-task ownership transfer.

Key Takeaways¶

Stop at the judgment threshold, not at failure or 100% — the goal is handing off the nuanced tail, not the whole task
Distilled summaries, not raw conversation history — compress what matters, discard the rest
Draft PRs and progress files are the durable handoff artifacts; the human should be able to pick up from git alone
Tool parity reduces translation friction: human and agent use the same interfaces
Progress visibility is a prerequisite — the human needs a signal to know when to engage