diff --git a/pkg/agent/checkpoint.go b/pkg/agent/checkpoint.go
index 952628e88..850ac4845 100644
--- a/pkg/agent/checkpoint.go
+++ b/pkg/agent/checkpoint.go
@@ -35,7 +35,7 @@ type (
 		// Approval-interrupted checkpoints carry pending tool calls.
 		PendingToolCalls []llm.ToolCall            `json:"pending_tool_calls,omitempty"`
 		PendingApprovals []llm.ToolCall            `json:"pending_approvals,omitempty"`
-		ApprovalInput    map[string]ApprovalResult `json:"approval_input,omitempty"`
+		ApprovalInput    map[string]ApprovalResult `json:"approval_input,omitempty"` // keyed by tool call ID
 
 		// Nested agent-as-tool suspension: one entry per suspended inner agent.
 		AllToolCalls     []llm.ToolCall         `json:"all_tool_calls,omitempty"`
diff --git a/pkg/agent/restore.go b/pkg/agent/restore.go
index 665e010b3..b51fbaa22 100644
--- a/pkg/agent/restore.go
+++ b/pkg/agent/restore.go
@@ -60,9 +60,6 @@ func restoreCheckpoint(
 	runID string,
 	registry AgentRegistry,
 ) (*Result, error) {
-	if cp.Version != CheckpointVersion {
-		return nil, fmt.Errorf("cannot restore: unsupported checkpoint version %d", cp.Version)
-	}
 	emitHook(agent, func(h RunHooks) { h.OnRunRestore(ctx, agent, cp) })
 
 	switch cp.Status {
diff --git a/pkg/agent/run.go b/pkg/agent/run.go
index 84b8eb363..5f7b5303b 100644
--- a/pkg/agent/run.go
+++ b/pkg/agent/run.go
@@ -374,6 +374,10 @@ func coreLoop(ctx context.Context, startAgent *Agent, inputMessages []llm.Messag
 			return s.finishRun(ctx, nil, fmt.Errorf("cannot complete: %w", err))
 		}
 
+		if s.turns >= s.agent.maxTurns {
+			return s.finishRun(ctx, nil, &MaxTurnsExceededError{MaxTurns: s.agent.maxTurns})
+		}
+
 		if ch := stopSignalFrom(ctx); ch != nil {
 			select {
 			case <-ch:
@@ -383,6 +387,7 @@ func coreLoop(ctx context.Context, startAgent *Agent, inputMessages []llm.Messag
 				if s.opts.checkpointStore != nil && s.opts.runID != "" {
 					if saveErr := s.opts.checkpointStore.Save(ctx, s.opts.runID, cp); saveErr != nil {
 						s.logger.ErrorCtx(ctx, "cannot save suspension checkpoint", log.Error(saveErr))
+						se.Checkpoint = cp
 					}
 				} else {
 					se.Checkpoint = cp
@@ -393,10 +398,6 @@ func coreLoop(ctx context.Context, startAgent *Agent, inputMessages []llm.Messag
 			}
 		}
 
-		if s.turns >= s.agent.maxTurns {
-			return s.finishRun(ctx, nil, &MaxTurnsExceededError{MaxTurns: s.agent.maxTurns})
-		}
-
 		fullMessages := buildFullMessages(s.systemPrompt, s.messages)
 
 		var responseFormat *llm.ResponseFormat
diff --git a/pkg/probo/agent_run_supervisor.go b/pkg/probo/agent_run_supervisor.go
new file mode 100644
index 000000000..ae4ec7f64
--- /dev/null
+++ b/pkg/probo/agent_run_supervisor.go
@@ -0,0 +1,416 @@
+// Copyright (c) 2026 Probo Inc <hello@getprobo.com>.
+//
+// Permission to use, copy, modify, and/or distribute this software for any
+// purpose with or without fee is hereby granted, provided that the above
+// copyright notice and this permission notice appear in all copies.
+//
+// THE SOFTWARE IS PROVIDED "AS IS" AND THE AUTHOR DISCLAIMS ALL WARRANTIES WITH
+// REGARD TO THIS SOFTWARE INCLUDING ALL IMPLIED WARRANTIES OF MERCHANTABILITY
+// AND FITNESS. IN NO EVENT SHALL THE AUTHOR BE LIABLE FOR ANY SPECIAL, DIRECT,
+// INDIRECT, OR CONSEQUENTIAL DAMAGES OR ANY DAMAGES WHATSOEVER RESULTING FROM
+// LOSS OF USE, DATA OR PROFITS, WHETHER IN AN ACTION OF CONTRACT, NEGLIGENCE OR
+// OTHER TORTIOUS ACTION, ARISING OUT OF OR IN CONNECTION WITH THE USE OR
+// PERFORMANCE OF THIS SOFTWARE.
+
+package probo
+
+import (
+	"context"
+	"encoding/json"
+	"errors"
+	"fmt"
+	"sync"
+	"time"
+
+	"go.gearno.de/crypto/uuid"
+	"go.gearno.de/kit/log"
+	"go.gearno.de/kit/pg"
+	"go.probo.inc/probo/pkg/agent"
+	"go.probo.inc/probo/pkg/coredata"
+	"go.probo.inc/probo/pkg/llm"
+)
+
+type (
+	AgentRunSupervisor struct {
+		pg             *pg.Client
+		store          *coredata.PGCheckpointStore
+		registry       agent.AgentRegistry
+		logger         *log.Logger
+		interval       time.Duration
+		leaseDuration  time.Duration
+		maxConcurrency int
+		workerID       string
+
+		mu      sync.Mutex
+		running map[string]*runHandle // runID -> handle
+	}
+
+	// runHandle wraps a stop channel with a sync.Once to prevent double-close panics.
+	runHandle struct {
+		stopCh chan struct{}
+		once   sync.Once
+	}
+
+	AgentRunSupervisorOption func(*AgentRunSupervisor)
+)
+
+var (
+	ErrAgentRunHeartbeatFailed = errors.New("agent run heartbeat failed")
+	ErrAgentRunLeaseLost       = errors.New("agent run lease lost")
+)
+
+func (h *runHandle) stop() {
+	h.once.Do(func() { close(h.stopCh) })
+}
+
+func WithAgentRunSupervisorInterval(d time.Duration) AgentRunSupervisorOption {
+	return func(s *AgentRunSupervisor) {
+		if d > 0 {
+			s.interval = d
+		}
+	}
+}
+
+func WithAgentRunSupervisorLeaseDuration(d time.Duration) AgentRunSupervisorOption {
+	return func(s *AgentRunSupervisor) {
+		if d > 0 {
+			s.leaseDuration = d
+		}
+	}
+}
+
+func WithAgentRunSupervisorMaxConcurrency(n int) AgentRunSupervisorOption {
+	return func(s *AgentRunSupervisor) {
+		if n > 0 {
+			s.maxConcurrency = n
+		}
+	}
+}
+
+func NewAgentRunSupervisor(
+	pgClient *pg.Client,
+	store *coredata.PGCheckpointStore,
+	registry agent.AgentRegistry,
+	logger *log.Logger,
+	opts ...AgentRunSupervisorOption,
+) *AgentRunSupervisor {
+	s := &AgentRunSupervisor{
+		pg:             pgClient,
+		store:          store,
+		registry:       registry,
+		logger:         logger,
+		interval:       10 * time.Second,
+		leaseDuration:  5 * time.Minute,
+		maxConcurrency: 5,
+		workerID:       uuid.MustNewV4().String(),
+		running:        make(map[string]*runHandle),
+	}
+
+	for _, opt := range opts {
+		opt(s)
+	}
+
+	return s
+}
+
+func (s *AgentRunSupervisor) Run(ctx context.Context) error {
+	var (
+		wg     sync.WaitGroup
+		sem    = make(chan struct{}, s.maxConcurrency)
+		ticker = time.NewTicker(s.interval)
+	)
+	defer ticker.Stop()
+	defer wg.Wait()
+	defer s.stopAll()
+
+	for {
+		select {
+		case <-ctx.Done():
+			return ctx.Err()
+		case <-ticker.C:
+			nonCancelableCtx := context.WithoutCancel(ctx)
+			s.recoverStaleRuns(nonCancelableCtx)
+			s.checkStopRequests(nonCancelableCtx)
+
+			for {
+				if err := s.processNext(ctx, sem, &wg); err != nil {
+					if !errors.Is(err, coredata.ErrResourceNotFound) {
+						s.logger.ErrorCtx(nonCancelableCtx, "cannot claim agent run", log.Error(err))
+					}
+					break
+				}
+			}
+		}
+	}
+}
+
+func (s *AgentRunSupervisor) processNext(
+	ctx context.Context,
+	sem chan struct{},
+	wg *sync.WaitGroup,
+) error {
+	select {
+	case sem <- struct{}{}:
+	case <-ctx.Done():
+		return ctx.Err()
+	}
+
+	var (
+		run              = coredata.AgentRun{}
+		now              = time.Now()
+		leaseOwner       = s.workerID
+		leaseExpiresAt   = now.Add(s.leaseDuration)
+		nonCancelableCtx = context.WithoutCancel(ctx)
+	)
+
+	if err := s.pg.WithTx(
+		nonCancelableCtx,
+		func(ctx context.Context, tx pg.Tx) error {
+			if err := run.LoadNextPendingForUpdateSkipLocked(ctx, tx); err != nil {
+				return err
+			}
+
+			run.Status = coredata.AgentRunStatusRunning
+			run.StartedAt = &now
+			run.LeaseOwner = &leaseOwner
+			run.LeaseExpiresAt = &leaseExpiresAt
+			run.UpdatedAt = now
+
+			if err := run.Update(ctx, tx, coredata.NewNoScope()); err != nil {
+				return fmt.Errorf("cannot update agent run: %w", err)
+			}
+
+			return nil
+		},
+	); err != nil {
+		<-sem
+		return err
+	}
+
+	handle := &runHandle{stopCh: make(chan struct{})}
+
+	s.mu.Lock()
+	s.running[run.ID.String()] = handle
+	s.mu.Unlock()
+
+	wg.Add(1)
+	go func(run coredata.AgentRun) {
+		defer wg.Done()
+		defer func() { <-sem }()
+		defer func() {
+			s.mu.Lock()
+			delete(s.running, run.ID.String())
+			s.mu.Unlock()
+		}()
+
+		runCtx, cancelRun := context.WithCancelCause(nonCancelableCtx)
+		defer cancelRun(nil)
+
+		heartbeatCtx, cancelHeartbeat := context.WithCancel(nonCancelableCtx)
+		defer cancelHeartbeat()
+		go s.heartbeatLease(heartbeatCtx, run.ID.String(), cancelRun)
+
+		runCtx = agent.WithStopSignal(runCtx, handle.stopCh)
+		s.executeRun(runCtx, &run)
+	}(run)
+
+	return nil
+}
+
+func (s *AgentRunSupervisor) heartbeatLease(
+	ctx context.Context,
+	runID string,
+	cancelRun context.CancelCauseFunc,
+) {
+	ticker := time.NewTicker(s.leaseDuration / 3)
+	defer ticker.Stop()
+
+	for {
+		select {
+		case <-ctx.Done():
+			return
+		case <-ticker.C:
+			expiresAt := time.Now().Add(s.leaseDuration)
+
+			if err := s.pg.WithConn(
+				ctx,
+				func(ctx context.Context, conn pg.Querier) error {
+					rowsAffected, err := coredata.HeartbeatAgentRunLease(ctx, conn, runID, s.workerID, expiresAt)
+					if err != nil {
+						return err
+					}
+
+					if rowsAffected == 0 {
+						return ErrAgentRunLeaseLost
+					}
+
+					return nil
+				},
+			); err != nil {
+				s.logger.ErrorCtx(ctx, "cannot heartbeat agent run lease", log.Error(err))
+				if errors.Is(err, ErrAgentRunLeaseLost) {
+					cancelRun(ErrAgentRunLeaseLost)
+				} else {
+					cancelRun(fmt.Errorf("%w: %w", ErrAgentRunHeartbeatFailed, err))
+				}
+				return
+			}
+		}
+	}
+}
+
+func (s *AgentRunSupervisor) executeRun(ctx context.Context, run *coredata.AgentRun) {
+	runID := run.ID.String()
+
+	var (
+		result *agent.Result
+		runErr error
+	)
+
+	if run.Checkpoint != nil {
+		// Resume from checkpoint.
+		s.logger.InfoCtx(ctx, "resuming agent run", log.String("run_id", runID))
+		result, runErr = agent.Restore(ctx, s.store, runID, s.registry)
+	} else {
+		// Start fresh.
+		s.logger.InfoCtx(ctx, "starting agent run", log.String("run_id", runID))
+
+		a, err := s.registry.Agent(run.StartAgentName)
+		if err != nil {
+			runErr = fmt.Errorf("cannot resolve agent %q: %w", run.StartAgentName, err)
+		} else {
+			var inputMsgs []llm.Message
+			if err := json.Unmarshal(run.InputMessages, &inputMsgs); err != nil {
+				runErr = fmt.Errorf("cannot unmarshal input messages: %w", err)
+			} else {
+				result, runErr = a.RunWithOpts(
+					ctx,
+					inputMsgs,
+					agent.WithCheckpointStore(s.store, runID),
+				)
+			}
+		}
+	}
+
+	// Heartbeat loss: another worker may have taken over. Do not commit
+	// any status — stale recovery will handle the row.
+	if cause := context.Cause(ctx); errors.Is(cause, ErrAgentRunLeaseLost) || errors.Is(cause, ErrAgentRunHeartbeatFailed) {
+		s.logger.WarnCtx(
+			context.WithoutCancel(ctx),
+			"agent run stopped after heartbeat failure; leaving status for stale recovery",
+			log.String("run_id", runID),
+			log.Error(cause),
+		)
+		return
+	}
+
+	// Infrastructure-triggered suspension (graceful shutdown): leave the
+	// row as RUNNING so stale recovery resets it to PENDING on restart.
+	// The checkpoint was already saved by coreLoop before returning
+	// SuspendedError, so Restore will pick up where it left off.
+	if runErr != nil {
+		if _, ok := errors.AsType[*agent.SuspendedError](runErr); ok {
+			s.logger.InfoCtx(
+				context.WithoutCancel(ctx),
+				"agent run suspended by infrastructure; leaving for stale recovery",
+				log.String("run_id", runID),
+			)
+			return
+		}
+	}
+
+	// Update run status based on outcome.
+	now := time.Now()
+	run.UpdatedAt = now
+	run.StartedAt = nil
+	run.LeaseOwner = nil
+	run.LeaseExpiresAt = nil
+
+	switch {
+	case runErr == nil:
+		run.Status = coredata.AgentRunStatusCompleted
+		if result != nil {
+			data, err := json.Marshal(result)
+			if err != nil {
+				s.logger.ErrorCtx(ctx, "cannot marshal agent run result", log.Error(err))
+			} else {
+				run.Result = data
+			}
+		}
+		run.StopRequested = false
+
+	default:
+		run.Status = coredata.AgentRunStatusFailed
+		msg := runErr.Error()
+		run.ErrorMessage = &msg
+	}
+
+	commitCtx := context.WithoutCancel(ctx)
+
+	if err := s.pg.WithTx(
+		commitCtx,
+		func(ctx context.Context, tx pg.Tx) error {
+			if err := run.Update(ctx, tx, coredata.NewNoScope()); err != nil {
+				return err
+			}
+
+			if run.Status == coredata.AgentRunStatusCompleted {
+				if err := run.ClearCheckpoint(ctx, tx, coredata.NewNoScope()); err != nil {
+					return err
+				}
+			}
+
+			return nil
+		},
+	); err != nil {
+		s.logger.ErrorCtx(commitCtx, "cannot commit agent run status", log.Error(err))
+	}
+}
+
+func (s *AgentRunSupervisor) checkStopRequests(ctx context.Context) {
+	var ids []string
+
+	if err := s.pg.WithConn(
+		ctx,
+		func(ctx context.Context, conn pg.Querier) error {
+			var err error
+			ids, err = coredata.LoadRunningStopRequestedIDs(ctx, conn)
+			return err
+		},
+	); err != nil {
+		s.logger.ErrorCtx(ctx, "cannot check stop requests", log.Error(err))
+		return
+	}
+
+	for _, id := range ids {
+		s.mu.Lock()
+		if h, ok := s.running[id]; ok {
+			h.stop()
+		}
+		s.mu.Unlock()
+	}
+}
+
+func (s *AgentRunSupervisor) stopAll() {
+	s.mu.Lock()
+	defer s.mu.Unlock()
+
+	for _, h := range s.running {
+		h.stop()
+	}
+}
+
+func (s *AgentRunSupervisor) recoverStaleRuns(ctx context.Context) {
+	if err := s.pg.WithConn(
+		ctx,
+		func(ctx context.Context, conn pg.Querier) error {
+			if err := coredata.ResetStaleAgentRuns(ctx, conn); err != nil {
+				return fmt.Errorf("cannot reset stale agent runs: %w", err)
+			}
+
+			return nil
+		},
+	); err != nil {
+		s.logger.ErrorCtx(ctx, "cannot recover stale agent runs", log.Error(err))
+	}
+}